Regular & Irregular Verbs PDF

Universität Potsdam
Helena Trompelt
Production of regular and

non-regular verbs
Evidence for a lexical entry complexity account
Spektrum Patholinguistik – Schriften | 2

Helena Trompelt
Production of regular and non-regular verbs

Evidence for a lexical entry complexity account
Universitätsverlag Potsdam
Bibliografische Information der Deutschen Nationalbibliothek
Die Deutsche Nationalbibliothek verzeichnet diese Publikation in der
Deutschen Nationalbibliografie; detaillierte bibliografische Daten sind im Internet
über http://dnb.d‐nb.de abrufbar.
Universitätsverlag Potsdam 2010

http://info.ub.uni‐potsdam.de/verlag.htm
Universitätsverlag Potsdam, Am Neuen Palais 10, 14469 Potsdam

Tel.: +49 (0)331 977 4623 / Fax: 3474
E‐Mail: verlag@uni‐potsdam.de
Die Schriftenreihe Spektrum Patholinguistik – Schriften wird herausgegeben

vom Verband für Patholinguistik e. V.
Das Manuskript ist urheberrechtlich geschützt.

Umschlagfotos:
Johannes Heuckeroth, http://www.flickr.com/photos/pfn/2682132140/
http://pfnphoto.com/
Kamil Piaskowski, http://mommus.deviantart.com/gallery/#/d20k31l
Satz: Martin Anselm Meyerhoff

Druck: docupoint GmbH Magdeburg
Zugl.: Potsdam, Univ., Diss., 2010

1st reviewer: Prof. Dr. Ria De Bleser
2nd reviewer: Prof. Dr. Thomas Pechmann
Day of submission: October 13, 2009
Day of oral defense: April 12, 2010
ISSN (print) 1869‐3822

ISSN (online) 1869‐3830
ISBN 978-3-86956-061-8
Zugleich online veröffentlicht

auf dem Publikationsserver der Universität Potsdam
URL http://pub.ub.uni‐potsdam.de/volltexte/2010/4212/
URN urn:nbn:de:kobv:517‐opus‐42120
http://nbn‐resolving.de/urn:nbn:de:kobv:517‐opus‐42120
Acknowledgement
Many thanks to all the people who have helped me both
personally and professionally to accomplish the work put forth in
this dissertation.
First and foremost, I am grateful to Prof. Dr. Thomas Pechmann

for his knowledgeable supervision and for always demanding
maximal clarity and accuracy of exposition.
This work would not have been possible without Prof. Dr. Ria De
Bleser. Her comments and discussions along the way were
important for the progress of this work.
The Graduiertenkolleg Universalität und Diversität at the

University of Leipzig and the Graduate Programme for Experimental
and Clinical Linguistics at the University of Potsdam supported me
in investigating an exciting phenomenon of German language
production. I was not only provided with financial support, but
also benefited from the contributions of all my remarkable
colleagues.
I am indebted to Dr. habil. Denisa Bordag for her constant and

close supervision. She made difficult things look natural and easy
and helped enormously by introducing me to the methods and
technical work. Special thanks for our extended and substantial
discussions of linguistic concepts too!
I would like to thank my friends Lars Meyer, Judith Heide, Tyko

Dirksmeyer, Kristina Kasparian and Antje Lorenz very much for
i
ii ACKNOWLEDGEMENT
smart comments on a first version of this thesis and for

thoroughly smoothing transitions.
Most important of all: I count myself truly fortunate for having

made and kept many dear friends, both before and during
graduate school – who made this time enriching and memorable.
By relating your problems and experiences to me, you helped me
to solve my own problems, and ultimately helped me to come to a
better.
Finally, the greatest debts of gratitude by far are to my parents

and my sister Antonia.
Contents
Acknowledgement ................................................................... i
Contents................................................................................iii
List of Tables........................................................................ vii
List of Figures........................................................................ix
Abbreviations..........................................................................x
0 Introduction ................................................................... 1
1 Regular and non-regular inflection................................... 5
1.1 Inflectional categories .......................................................... 5

1.2 Paradigms and classes ......................................................... 7
1.3 Language typology .............................................................. 9
1.3.1 German verbal inflectional system .................................. 10
1.3.2 Comparison of English and German inflectional system... 11
1.4 Aspects of regular and non-regular nominal inflection ........ 13
1.5 Summary........................................................................... 17
2 Approaches to regular and non-regular inflection ............19
2.1 Articulation latencies of regular and non-regular verbs ....... 19

2.2 Dual Route models of language production ........................ 21
2.3 The Words and Rules Theory ............................................ 22
2.3.1 The blocking mechanism................................................ 24
iii
iv CONTENTS
2.3.2 Psycholinguistic evidence for a regular/non-regular

dissociation of verbs....................................................... 27
2.3.3 Representation of regularity in the Words and Rules
Theory........................................................................... 35
2.3.4 The Words-and-Rules-Theory’s difficulties...................... 37
2.4 Connectionist accounts ...................................................... 38
2.4.1 The Pattern Associator .................................................. 40
2.4.2 Strengths and weaknesses of connectionist models ......... 44
2.5 Summary........................................................................... 45
3 Psycholinguistic models of language production ..............47
3.1 Lexical access and lexical selection..................................... 47

3.2 The Levelt Model (Levelt, 1999)........................................ 53
3.2.1 Architecture................................................................... 53
3.2.2 Diacritic parameters....................................................... 56
3.3 The Interactive Activation Model....................................... 58
3.4 The Independent Network Model....................................... 61
3.5 Discrete versus cascaded processing ................................... 65
3.6 Remarks on diversity of models.......................................... 65
3.7 Producing morphologically complex words ......................... 66
3.8 Morphological processing in comprehension ....................... 68
4 Representation and processing of grammatical features ..71
4.1 Representation of linguistic information in the mental

lexicon............................................................................... 71
4.1.1 Structure of the mental lexicon ...................................... 72
4.1.2 Underspecified lexical entries.......................................... 74
4.2 Internal and external features ............................................ 75
4.3 Processing grammatical gender.......................................... 77
4.4 Processing declension and conjugation classes.................... 79
5 Tense............................................................................83
6 The empirical stance......................................................87

CONTENTS v
6.1 Why and how regularity might be represented ................... 87

6.2 The regularity congruency effect........................................ 89
7 Experiments ..................................................................95
7.1 Experiment 1 – Present tense ............................................ 95

7.1.1 Methods ........................................................................ 96
7.1.2 Results........................................................................... 99
7.1.3 Discussion.................................................................... 102
7.2 Experiment 2 – Past tense............................................... 103
7.2.1 Method........................................................................ 103
7.2.2 Results......................................................................... 104
7.2.3 Discussion.................................................................... 106
7.3 Experiment 3 – Present and past tense ............................ 108
7.3.1 Method........................................................................ 109
7.3.2 Results......................................................................... 109
7.3.3 Discussion.................................................................... 115
7.4 Non-regular verbs revisited .............................................. 117
7.5 Experiment 4................................................................... 118
7.5.1 Methods ...................................................................... 119
7.5.2 Results......................................................................... 121
7.5.3 Discussion.................................................................... 125
7.6 Discussion of Experiments 1-4 ......................................... 126
7.6.1 Critical evaluation of the picture-word interference
paradigm ..................................................................... 127
7.6.2 A caveat ...................................................................... 130
7.6.3 Intermediate conclusion................................................ 132
7.7 Experiment 5................................................................... 133
7.7.1 Method........................................................................ 134
7.7.2 Results......................................................................... 135
7.7.3 Discussion.................................................................... 137
8 General Discussion.......................................................141
References ..........................................................................157
vi CONTENTS
Appendix ............................................................................167
Stimuli for Experiment 1, 2 and 3................................................. 167

Stimuli for Experiment 4 and 5..................................................... 168
LIST OF TABLES vii
List of Tables
TABLE 1. PERCENTAGE OF MISSING VALUES (EXPERIMENT 1)......... 100
TABLE 2. MEAN RESPONSE LATENCIES AND ACCURACY PROPORTIONS
(EXPERIMENT 1).......................................................... 100
TABLE 3. MEAN RESPONSE LATENCIES, BY REGULARITY, SOA AND
DISTRACTOR (EXPERIMENT 1). ..................................... 101
TABLE 4. ACCURACY PROPORTIONS, BY REGULARITY AND
(EXPERIMENT 2).......................................................... 104
TABLE 7. MEAN RESPONSE LATENCIES, BY REGULARITY, SOA AND
TABLE 8. ACCURACY PROPORTIONS, BY REGULARITY AND
TABLE 10. MEAN RESPONSE LATENCIES AND ACCURACY
(EXPERIMENT 3).......................................................... 110
TABLE 11. MEAN RESPONSE LATENCIES, BY REGULARITY, SOA,
DISTRACTOR AND TENSE
(EXPERIMENT 3).......................................................... 111
TABLE 12. MEAN RESPONSE LATENCIES AND RESPONSE LATENCY
DIFFERENCE BETWEEN CONTROL
AND EXPERIMENTAL STIMULI (EXPERIMENT 3). .............. 112
TABLE 13. MEAN RESPONSE LATENCIES, BY REGULARITY AND TENSE
(EXPERIMENT 3).......................................................... 113
TABLE 14. MEAN RESPONSE LATENCIES, BY DISTRACTOR AND SOA
(EXPERIMENT 3).......................................................... 113
TABLE 15. MEAN RESPONSE LATENCIES, BY DISTRACTOR, TENSE AND
SOA (EXPERIMENT 3).................................................. 114
viii LIST OF TABLES
TABLE 16. ACCURACY PROPORTIONS, BY REGULARITY, TENSE AND

(EXPERIMENT 4).......................................................... 122
TABLE 19. MEAN RESPONSE LATENCIES, BY REGULARITY, SOA,
DISTRACTOR AND TENSE (EXPERIMENT 4). .................... 123
TABLE 20. MEAN RESPONSE LATENCIES AND RESPONSE LATENCY
DIFFERENCE BETWEEN CONTROL AND EXPERIMENTAL
STIMULI (EXPERIMENT 4). ............................................ 124
TABLE 21. MEAN RESPONSE LATENCIES, VARIED BY REGULARITY AND
TENSE (EXPERIMENT 4)................................................ 125
TABLE 23. MEAN RESPONSE LATENCIES, BY REGULARITY AND TENSE
(EXPERIMENT 5).......................................................... 136
TABLE 24. PERCENTAGE OF MISSING VALUES, BY TENSE AND
REGULARITY (EXPERIMENT 5). ..................................... 137
List of Figures
FIGURE 1. CLASSIFICATION OF GERMAN VERBS ................................ 11
FIGURE 2. AGREEMENT AND REGULARITY AS FACTORS FOR INFLECTION
IN GERMAN AND ENGLISH............................................... 11
FIGURE 3. LANGUAGE PRODUCTION MODEL .................................... 54
FIGURE 4. STRUCTURED LEXICAL ENTRY FOR A NON-REGULAR VERB .. 75
FIGURE 5. REPRESENTATION OF LEXICAL INFORMATION AS GENERIC
NODES.......................................................................... 78
FIGURE 6. STIMULI EXAMPLES FOR THE VERB TRINKEN..................... 97
FIGURE 7. SIMPLE AND COMPLEX LEXICAL ENTRIES OF GERMAN VERBS. 147
FIGURE 8. LEXICAL ENTRY OF THE HYBRID GERMAN VERB SINGEN. .. 149
FIGURE 9. MODEL FOR THE GENERATION OF REGULAR AND
NON-REGULAR VERBS. .................................................. 153
ix
Abbreviations
AAM Augmented Addressed Model
(Caramazza et al., 1988)
DRM Dual Route Model
ERP Event Related Potential
Gr. Greek
hyb hybrid
IA Model Interactive Activation Model (Dell, 1986)
IN Model Independent Network Model (Caramazza, 1997)
irr irregular
LD Lexical Decision
ms milliseconds
nreg non‐regular
PET Positron Emission Tomography
PWI Picture‐Word Interference
reg regular
RT Reaction Time
SMA Supplementary motor area
SOA Stimulus Onset Asynchrony
WR Words and Rules Theory (Pinker, 1999)
x
0 Introduction
The incredible productivity and creativity of language depends on
two fundamental resources: a mental lexicon and a mental grammar
(Chomsky, 1995; Pinker, 1994). The mental lexicon stores
information and masters the arbitrariness of the language. It is a
repository of idiosyncratic and word specific, i.e. atomic non‐
decomposable information. For example, the mental lexicon
contains the arbitrary sound‐meaning pair for dog and the
information that it is a noun. The mental lexicon also comprises
complex idiosyncratic phrases such as It rains cats and dogs, the
meaning of which cannot be derived transparently from the
constituents (Swinney & Cutler, 1979; Gibbs, 1980).
In addition to the mental lexicon, language is also made up of
rules of grammar constraining the computation of complex
expressions. These rules of grammar enable human speakers to
produce and understand sentences that they have not encountered
before. The meaning, then, can be derived from the constituents
and knowledge about rules. Not only do these determine
sequential ordering of constituents but also hierarchical relations.
The recipient of the message The dog tammed the crig knows that
the dog is the actor of a past action and that the dog did
something to the entity crig. To compute an infinite number of
new structures from stored elements and to derive their meaning
is an enormous grammatical ability and the source of productivity
and creativity of human language (Chomsky, 1995).
The computational component of the language faculty can be
found at various levels in natural languages: e.g. at the sentence
level (syntax) and the level of complex words (morphology). Rules
manipulate meaning and structure of symbolic representations.
Applying recursively, a limited set of units and rules is the core
for combinations of unlimited number and unlimited length.
1
2 CHAPTER 0: INTRODUCTION
Moreover, rules apply on abstract symbolic categories rather than

on particular words. They can generate unusual combinations
(colourless green ideas) and nonsense sentences (The dog tammed the
crig).
This dichotomy of the mental lexicon on the one hand and the
mental grammar on the other hand can help us to better
understand the brain mechanisms we employ to process language.
Symbol manipulation underlies classification and identification of
classes of symbols suppressing irrelevant information and
drawing inferences that are likely to be true of every member of
the class, even individual members that have not previously been
encountered. The ability to handle an unfamiliar symbol as a
member of a class is central to cognition. Hence, the resolution of
this dichotomy in mental lexicon and mental grammar is not only
of interest to psycholinguistics but to psychology, neurosciences,
linguistics and artificial intelligence as well. Ullman (2001a; 2001b)
generalises that the distinction between stored and computed
representations in language is tied to two distinct brain memory
systems: declarative and procedural memory. The declarative
memory system is devoted to learning and remembering facts and
events, whereas the procedural system is responsible for
sequencing of representations or motor actions.
The concepts of mental lexicon and mental grammar have been
thoroughly tested in the context of the use of regular versus non‐
regular inflectional morphology. Inflection is one way languages
express grammatical relations by changing the form of a word to
give it extra meaning. The inflection of verbs encapsulates the
issues of lexicon and grammar. Regular verbs (walk‐walked;
lachen‐lachte [to laugh‐laughed]) are computed by a suffixation
rule in a neural system for grammatical processing; irregular verbs
(run‐ran) are retrieved from an associative memory. A heated and
polarizing, though fruitful, debate concerns the processing and
representation of regular and non‐regular verb forms (Marcus,
Brinkmann, Clahsen, Wiese & Pinker, 1995; Pinker, 1997; Clahsen,
1999).
3
„Perhaps regular verbs can become the fruit flies of the neuroscience of
language – their recombining units are easy to extract and visualize
and they are well studied, small and easy to breed.“ (Pinker, 1997)
The comparison by Steven Pinker of regular verbs being the fruit

flies of neuroscience of language highlights their importance and
potential. Their nature is as fascinating as the genome of those
little flies with their protuberant eyes, certain to appear every
summer. The study of regular verbs is supposed to uncover
meaningful evidence about human cognition like drosophila did
for genetics.
Out of the above mentioned debate two approaches have
emerged: one camp assumes associative memory mechanisms
(Rumelhart, McClelland & the PDP research group, 1986;
Ramscar, 2002; MacWhinney & Leinbach, 1991; Daugherty &
Seidenberg, 1994), while the other camp presumes the existence of
linguistic rules (Berko, 1958; Heidolph, Fläming & Motsch, 1981;
Eisenberg, 1994). Pinker (1991; 1999) finally combined both
approaches to the Dual Route Model.
However, little research has been devoted to the production of
regular and non‐regular verbs and even less to aspects of tense.
The aims of this thesis are to explore the cognitive reality of
regularity1, its representation and inflectional processes involved
in the production of regular and non‐regular verbs in unimpaired
speakers. Chapters 1‐5 provide the theoretical background
(empirical findings and psycholinguistic models) as well as
morphological concepts of linguistic theory. Linguistic factors
play a crucial role in psycholinguistic processing and must be
closely considered in motivating hypotheses and modelling.
Chapter 6 and 7 summarize and discuss data of four picture‐
word‐interference experiments exploring the cognitive status of
regularity as well as a picture naming experiment which validates
1 The term regularity includes non‐regularity. It is a superordinate concept and

refers to the general property of items to be regular or not. The non‐existence of the
concept/term indicates that it has not been analysed as general feature before but
only with its specific values regular/non‐regular/irregular. Regularity means more
or less “regularity status”.
4 CHAPTER 0: INTRODUCTION
the different morphological inflectional processes assumed and

devoted to lexicon and grammar. The discussion in chapter 8 is a
re‐consideration and evaluation of the current models on verb
processing in the light of the experimental data.
1 Regular and non-regular
inflection
Theories of language processing often draw upon the accounts
made by theoretical linguistics. Theoretical linguistics seeks to
describe mental representations. The linguist’s morphological
decomposition of words has often been examined by
psycholinguists asking whether there is any correspondence
between linguistic theory and the way morphologically complex
words are segmented by the speaker or hearer during online
production and comprehension. The first chapter aims at
anchoring the psycholinguistic question studied in linguistic
theory.
Like in most Germanic languages, German verbs can be
classified into two fundamentally different classes: regular (weak)
and non‐regular (strong) verbs. The crucial difference between
these classes lies in the formation of their past tense stem.
Historically, non‐regular verbs are the base stock of all Germanic
verbs. Nowadays, non‐regular verbs constitute a minority of all
verbs as many were regularised with language change.
1.1 Inflectional categories
Inflection is one way languages express grammatical relations.

Inflectional processes are specific for certain parts of speech. For
German verbs, inherent2 inflectional categories include person,
tense, number, voice and mood (cf. Spencer & Zwicky, 1998; Bickel &
2 A non‐inherent inflectional category of verbal inflection is for example subject‐

verb‐agreement.
5
6 CHAPTER 1: REGULAR AND NON‐REGULAR INFLECTION
Nichols, 2007). Typologically, different languages may also specify

gender, honorificity, animacy or definiteness on the verb. German
nouns are inflected for number and case. Inflectional operations
often are systematic and exceptionless.
Regularity is a crucial property to create a tense marked stem
of a verb. Regularity, however, is inherent to a verb’s grammatical
properties because it is arbitrary. The verbs of Indo‐European
languages generally have distinct suffixes to specify present and
past tenses. Verb inflection in German is usually divided into two
parts: regular (weak) and non-regular (strong) inflection
(conjugation). Regular verbs (e.g. (1) spielen [to play]) have only
one stem and take regular affixes in both past (–te) and present (–
t). Non‐regular verbs like (2) blasen [to blow] have alternating
stems.
(1) ich spiel‐e, er spiel –t, er spiel‐te [I play, he play‐s, he play‐ed]

(2) ich blas‐e, er bläs‐t, er blies- [I blow, he blow‐s, he blew]
The preparation of the stem and the inflection of the verb form for
person and number are two distinct processes. Person and
number agreement is formed independently of the verb’s
regularity without any exceptions (disregarding suppletives). This
means that person suffixes (e.g. –st for second person singular) are
added to the tense marked stem, be it regular (du lachte-st) or non‐
regular (du trank-st) (Clahsen, 1996; Clahsen, Eisenbeiss, Hadler &
Sonnenstuhl, 2001). The primary mechanism in verb production is
the generation of the stem, person and number inflection follows
only if the stem is completed.
Most inflectional categories fulfil requirements of syntax.
Inflection can be governed syntactically realising, for instance,
agreement relations. Then inflection is occasioned by a governing
head in a sentence and is highly productive. (cf. he and -s in (3a)).
Also determiners and adjectives in many languages agree with the
head noun as in the German example in (3b).
1.2 PARADIGMS AND CLASSES 7
(3) a. he sing-s/we sing

b. ein knusprig-es Brot
a-NOM.SG.NEUTR crispy-NOM.SG.NEUTR bread-
NOM.SG.NEUTR
However, Bickel and Nichols (2007) point out that the sentence
level is not the only dimension that conditions inflectional
processes:
“The relevant grammatical environment can be either syntactic or

morphological.” (Bickel & Nichols, 2007)
By contrast to syntactically driven inflectional processes, the

inflection can be determined by lexical entries’ inherent
properties, i.e. morphological properties like aspect in Russian.
Aspect in Russian classifies verbal bases into perfective and
imperfective in a highly non‐regular manner. The future tense of
imperfective verbs is formed analytically (periphrastically), but
synthetically (by means of a morphological process) if the verb is
perfective. Here, inflection is determined by lexical aspect. Also
the inflection of verbs in German depends on the inherent
regularity status of the verbs.
1.2 Paradigms and classes
The term paradigm is derived from the Greek word for pattern (Gr.
παράδειγμα parádeigma). It contains the full set of inflectional
endings a stem can have. With respect to verbal inflection, it
provides the material to instantiate all possible word forms of the
verb and serves as a model or example for all others. Paradigms
are often set out as a learning tool in language teaching. By
knowing the pattern a new word belongs to, one can easily inflect
the new word by generalising on the basis of known words of the
same paradigm ‐ usually by adding a suffix to indicate a change in
its grammatical function. For German, see example (4) for

paradigmatic relations with respect to inflection of verbs for
person in present tense (a+b) and past tense (c+d).
(4) a. ich lach‐e, du lach‐st, er lach‐t…

[I laugh, you laugh, he laugh‐s…]
b. ich geh‐e, du geh‐st, er geh‐t…
[I go, you go, he go‐(e)s…]
c. ich lach‐te-ø, du lach‐te‐st, er lach‐te‐ø…
[I laughed, you laughed, he laughed…]
d. ich ging‐ø, du ging‐st, er ging‐ø…
[I went, you went, he went…]
The person/number suffixes for conjugating lachen and gehen are

identical but the behaviour of the stem renders them distinct. A
multi‐paradigm analysis is necessary as there is no way to predict
whether a verb employs allomorphic stems or not. The term
paradigm is used in various senses (for a review see Carstairs‐
McCarthy, 2000). I adhere to the most common usage of the term
paradigm according to which paradigms correspond to the
taxonomy of regular and non-regular verbs (cf. (Carstairs, 1987).
Hence, geben (to give) and singen [to sing] belong to the same
paradigm; gehen (to go) and lachen [to laugh] belong to two different
paradigms: the latter forms its past tense by adding –te to the stem
and the former has a special past tense stem.
Recurring regularities within the German non‐regular verbs
are conjugation or inflectional classes (Ablautreihen). Ablaut refers
to an internal modification involving vowel change. For instance,
singen - sang - gesungen [sing – sang – sung] is a case of Ablaut.
Inflectional classes stand out due to phonological predictability
(Wurzel, 1984). The verb trinken [to drink] belongs to the same
Ablaut pattern as singen (to sing) above as it is inflected alike:
trinken ‐ trank – getrunken [drink ‐ drank ‐ drunk] (Wiese, in
prep.). Hence, conjugation classes are subclasses of the non‐
regular paradigm like singen - sang - gesungen and trinken ‐ trank
– getrunken, which have the same vowel sequences. geben [to give]
1.3 LANGUAGE TYPOLOGY 9
and singen [to sing] belong to the non‐regular paradigm but to

different conjugation classes.
Conjugation classes reflect regularities among non‐regular
verbs. However, the empirical motivation of conjugation classes is
not without counter examples. The same phonological
environment is not necessarily a reliable predictor of the correct
past tense forms, e.g. the inflections of sink and think follow
different ablaut patterns (sank, sunk versus thought, thought) and
therefore belong to different phonological classes.
Conjugation classes are much more present in Romance
languages like French or Italian. Here, verbs which are
phonologically alike, i.e. having the same vowel pattern and/or
ending, belong to the same type of conjugation. Normatively and
for first (L1) and second language (L2) learning they are defined
as classes.
1.3 Language typology
Anglo‐American research constitutes a large part of the

psycholinguistic scientific work. Hence, English is studied in
much more detail than any other language. This section serves to
highlight the structural similarities of Indo‐European languages
but also their differences. Of most interest here, and therefore in
the spotlight, are English and German.
The Indo‐European languages comprise a family of several
hundred related languages. Indo refers to the Indian subcontinent,
since geographically the language group extends from Europe in
the west to India in the east. German and English are the most
widely spoken languages of the branch of Germanic languages
(Gordon, 2005). They belong to one family and are genetically
related.
Germanic languages share several attributes compared to other
families. The verbal system of tense and aspect has undergone
syncretism and has reduced to the present tense and the past tense
(the preterite). In English and German, a large set of verbs uses a
dental suffix ([d] or [t] in English and [tə] in German) instead of
vowel alternation to indicate past tense. These are known as

Germanic weak verbs and the remaining as Germanic strong verbs3.
1.3.1 German verbal inflectional system
The German verb system is indeed organised in a greater diversity

than shown so far and greater than for example the English one. It
is set up by three basic paradigms. The non-regular paradigm is
comprised of hybrid and irregular verbs. In addition to regular
verbs (e.g. (5) spielen [to play]), German has a second and third
type of verb which most authors group as irregular verbs
disregarding possible dissociations. Hybrid verbs are completely
regularly inflected in the present tense but employ irregular stems
in the past tense. Hence, they have also more than one stem (e.g.
(7) singen [to sing] – sang [sang]). Irregular verbs have irregular
forms in both the present and the past tense (e.g. (6) blasen [to
blow])4
(5) ich spiel‐e, er spiel –t, er spiel‐te [I play, he play‐s, he play‐ed]

(6) ich blas‐e, er bläs‐t, er blies- [I blow, he blow‐s, he blew]
(7) ich sing‐e, er sing‐t, er sang- [I sing, he sing‐s, he sang]
Hybrid verbs unify regular and irregular forms in one paradigm.

The crux of the matter: their typical and common characteristic are
stem vowels like /ei/ in beißen and /i:/ in biegen which do not have
fronted counterparts and therefore cannot mutate to umlaut. As in
the previous research, I will use the term non-regular to refer to
both hybrid and irregular verbs as opposed to regular verbs. The
whole picture is exemplified in Figure 1.
3 Weak and strong is a synonym terminology for regular and non‐regular.

Regarding processing aspects, regular and non‐regular obviously describe the
assumed cognitive processes more accurately.
4 In the DUDEN grammar (Kunkel‐Razum, 2006) another term “Mischverben” is
introduced which defines verbs with participles irregularly ending in ‐n, e.g.
mahlen: er mahlt, er mahlte, er hat gemahlen [to mill] although their past tense is
regular. Those are another potential object of investigation.
1.3 LANGUAGE TYPOLOGY 11
Figure 1. Classification of German verbs
1.3.2 Comparison of English and German inflectional system

The following section discusses two reasons why English is not
the most appropriate language (at least not the only one) to test
morphological dissociations between regular and non‐regular
verbs in different tenses. Of particular relevance to the following
empirical work, three dimensions will be discussed in order to
compare the inflectional systems of English and German:
Agreement, regularity and tense. In Figure 2, “x” indicates overt
phonological marking for person/number agreement and
regularity depending on tense as tense is in the focus of the
current study.
German English
Agreement Regularity Agreement Regularity
Present x x x –
Past x x – x
Figure 2. Agreement and Regularity as factors for inflection in German

and English.
It can be seen that in German, agreement is potentially visible on

word forms in both tenses (du gehst [you go] – du gingst [you
went]). English also inflects in the present tense for person and
number, although in a rather rudimentary way. Its overt inflection
occurs only for the third person singular. Person and number
agreement in the English past tense can only be derived from the
context. Verbs are not overtly marked for this feature (exception:
to be).
Consider the German Regularity column in Figure 2 above.
The crosses indicate overtly visible, regularity dependent
inflectional processes in the present tense as well as in the past
tense. Verbs like brechen [to break] are responsible for the cross in
the Regularity by Present cell. Some German verbs show non‐
regularities in the present tense (ich breche [I break] – er bricht [he
breaks]). However, the dissociation between regular and non‐
regular verb inflection in English targets only past tense
generation5. Regular verbs concatenate the stem and the suffix –ed
(laugh – laughed) – very similar to German (lachen – lachte). This is a
productive process which can be applied to novel verbs. The past
forms of the non-regular English verbs are morphologically simple
and correspond to past stems (take – took) ‐ insufficient, however,
to study the interplay of agreement and regularity. According to
most accounts, to produce non‐regular past tense verbs, English
speakers just have to retrieve the particular form from the mental
lexicon. On the contrary, the past tense forms of the German verbs
are morphologically complex. They consist of the past stem and an
inflectional suffix expressing person.
Languages tend to regularise inflection. The German verb
fragen [to ask] for example, was inflected non‐regularly in the past,
cf. er frug [he asked.IRR], which is out of use today. Nowadays,
the widely used form is er fragte [he asked.REG]. Non‐regular
inflection is not productive. New (i.e. loan) words in German and
other languages as well as novel words are conjugated regularly
(like chatten [to chat] er chattete, tammen [to tam] er tammte). The
same observations hold for participle constructions (Clahsen,
1996; Clahsen, Hadler & Weyerts, 2004). In English, there are also
doublets like to dream whose past tense can either be regular
dreamed, dreamed or non‐regular dreamt, dreamt.
5 Due to the suppletive forms of the English ‘to be’ and ‘to have’ these verbs are
sometimes also considered as showing “non‐regular” inflection in present tense.
However, the more general observation is that English verbs show no non‐regular
forms in present tense at all.
1.4 ASPECTS OF REGULAR AND NON‐REGULAR NOMINAL INFLECTION 13
Finally, although German and English belong to the West

Germanic language family, differences in representation and
processing of verbs between English and German could result from
the agreement asymmetry with regularity in English. Language‐
specific processing of regular and non‐regular verbs has been
shown by Meunier & Marslen‐Wilson (2004). They have examined
word‐level regular ‐ non‐regular verb contrasts in French, a
Romance language. In cross‐modal priming experiments, effects
seemed not to be a function of verb type: regular and non‐regular
verbs yielded the same amount of priming in French whereas
English non‐regular verbs showed reduced priming in the same
task as opposed to regular verbs. French regular and irregular
inflected forms do not show different priming patterns. The
facilitatory effects of morphologically related primes are equally
strong whether they involve the same or different underlying stem
as their targets. Pairs like buvons/boire [we drink/to drink] prime just
as well as pairs like aimons/aimer [we love/to love]. This pattern of
results was confirmed with the masked priming technique ruling
out the influence of semantic factors.
Though the French and English data were in contrast, French
and Italian results were compatible (for experiments on Italian see
Orsolini & Marslen‐Wilson, 1997). Compared to English, French
and Italian are Romance languages with richer inflection for
different types of tense, aspect and person suffixes. Irregular past‐
tense forms occur in a morphologically more structured and
phonologically more predictable linguistic environment than in
English. In connection with these and the upcoming experimental
findings it is important to point out once more that English and
German belong to the same, i.e. West Germanic, language family.
1.4 Aspects of regular and non-regular nominal

inflection
Although the focus of this study is on cognitive mechanisms of

regular and non‐regular verb processing, regularity is associated
with noun processing as well. This section will give a brief
introduction and comparison between regularity of nouns and

verbs. The comparison of nominal and verbal inflection can shed
light on the discussion of the morphological structure of verbs.
Nouns have been more extensively studied than verbs. A study by
Tylor, Bright, Fletcher and Stamatakis, (2004) suggests that noun
and verb stems do not differ regarding their representation, but to
generate morphologically complex verbs mostly those neural
systems are engaged which are involved in processes of morpho‐
phonology and syntax.
In English and German, nouns can show regularities in their
plural formation. In English, plural is mostly built by affixing –s to
the stem (8a), apart from exceptions like (8b). Exceptions
unpredictably employ allomorphs in the plural that have to be
learnt during language acquisition.
(8) a. chair – chairs [reg]

b. foot – feet [nreg]
In contrast to English, the German plural system is highly diverse.

Noun inflection in highly inflecting languages like German is
organised in declensional classes like verbs in conjugational
classes (Bordag & Pechmann, 2009, for Czech). German has five
plural affixes, -(e)n, -s, -e, -er and -∅ (see 9 (a‐e)). Example (9f)
additionally involves umlauting of the back stem vowel ‐ a
frequent and unpredictable phenomenon in German6.
(9) a. Frau – Frauen [nreg] [woman ‐ women]

b. Auto – Autos [reg] [car ‐ cars]
c. Hund – Hunde [nreg] [dog ‐ dogs]
d. Kind – Kinder [nreg] [child ‐ children]
e. Daumen ‐ Daumen [nreg]7 [thumb ‐ thumbs]
f. Haus ‐ Häuser [nreg] [house ‐ houses]
6 In fact, some researchers (e.g. Penke, 2006) assume umlauting to be rule‐driven

and therefore not to be an unpredictable process. Another argument pro
unpredictability is considered in example (10).
7 Non‐regularity of Daumen is questionable, because Computer as a loan word
exhibits the same null‐inflection [*Computer‐s].
1.4 ASPECTS OF REGULAR AND NON‐REGULAR NOMINAL INFLECTION 15
g. Erbse ‐ Erbsen [reg] [pea ‐ peas]

(cf. Clahsen, Rothweiler, Woest & Marcus, 1992)
Plural formation with –s (9b) is a productive process although –s is

a very infrequent marker. Marcus (1995), Marcus et al. (1995) and
Clahsen et al. (Clahsen, Eisenbeiss & Sonnenstuhl, 1997) argue for
s‐plurals to be a regular default process. –s plurals meet the
criterion of productivity in German and English as the plural
suffix –s is generalised to new and pseudowords as well as
acronyms etc., hence, probably by rule application (for a detailed
review see Clahsen, 1999). All other noun plurals are assumed to
be irregular and stored as fully inflected forms in the mental
lexicon in this Dual Route account.
However, there is a controversy according to the default
inflectional process of nouns. Penke & Krause (2002) investigated
whether other classes of noun plural formation can also be
regarded as regular. The most frequent plural marker is –n
(Sonnenstuhl & Huth, 2002). ‐n is completely predictable for
feminine nouns that end in schwa in the singular (9e) while –n is
not predictable for masculine or neuter nouns. Penke and Krause
(2002) presented data on German ‐n plurals that suggested that
feminine noun plurals on –n —although clearly not the German
default plural— result from a process of regular affixation.
Agrammatic aphasic patients were instructed to transform a given
singular noun into the respective plural form. These data
provided evidence that non‐feminine –n plurals and feminine –n
plurals can be affected differently in agrammatism. Moreover, this
dissociation is supported by a word form frequency effect for non‐
feminine n‐plural nouns. Further evidence comes from a lexical
decision (LD) experiment with unimpaired participants whose
performance showed a qualitative distinction between these two
types of ‐n plurals. Altogether, Penke & Krause (2002) concluded
that –n plurals cannot be uniformly treated. They proposed that
feminine nouns ending in schwa are built by regular affixation
like –s plural and others are stored as full forms.
It is of importance for correct morphological processing how
the word is inflected – either regularly or non‐regularly. In
German, the crucial point that causes unpredictability and non‐

regularity is not only the suffix but also the umlaut which may co‐
occur with -e, ‐er and -∅. Compare (10a) and (b):
(10) a. Hund – Hunde

b. Kuh – Kühe
Both are e‐plurals but Kühe employs a new stem and that class
therefore bears characteristics of non‐regular inflection. The
examples in (10) also demonstrate that it is not clearly predictable
whether the inflectional suffix or the stem itself is non‐regular.
This confound complicates the study of regularity of nouns.
The two systems – verb inflection and noun inflection – are not
strictly comparable anyway and neither are transferable
principles, because the types of inflectional processes are different.
Due to theoretical and formal differences between nouns and
verbs, it is widely assumed that both word classes have different
demands on cognitive processing (Kauschke, 2007; Druks, 2002;
Davidoff & Masterson, 1996).
Word classes can be dissociated by distinct connected
properties. Kauschke (2007) discusses criteria of several linguistic
areas. Basically, conceptual representations of nouns8 denote
objects and verbs denote actions. Clearly, verbs and nouns differ
in their inflectional categories. For most Indo‐European languages,
verbal inflection includes tense, mood, aspect and congruency
with the subject of a sentence. Critical for nominal inflection are,
however, definiteness, number and case. But there are not only
morphological criteria to mention; syntactically, verbs and nouns
vary in distribution, position, function and argument structure.
These descriptive analyses have often been confronted with the
cognitive dimensions. There is a consensus that verbs and nouns
differ in their cortical representations (Sahin, Pinker & Halgren,
2006). Many researchers (e.g. Laiacona & Caramazza, 2004) argue
for a double dissociation between the availability of nouns and
verbs in aphasia. There are brain damaged patients who show a
8 Except nominalisations like the destruction.

1.5 SUMMARY 17
selective difficulty in producing nouns in the context of a

relatively spared ability to produce verbs, while other patients
show the opposite pattern (Caramazza & Hillis, 1991).
In speech production, verbs are retrieved and articulated more
slowly than nouns (Szekely, DÀmico, Devescovi, Federmeier,
Herron, Iyer, Jacobsen & Bates, 2005). Szekely and colleagues
examined factors influencing noun and verb production in a
timed picture naming paradigm. Despite careful control of
multiple factors like age of acquisition, length, naming agreement
and complexity etc., massive differences [p.7] in noun and verb
production persisted. Action naming was shown to take
significantly longer than object naming and can then only be
attributed to word class.
All in all, it is debatable whether differences stem from
grammatical categories per se or from related and confounding
properties. But in most studies, the theoretical and the cognitive
view do not dissect. An exception is Siri et al. (2008).
1.5 Summary
This dissertation investigates the inflectional processes involved in

the production of regular and non‐regular German verbs, in
particular the generation of past and present tense stems. In
German, three types of verbs are known, namely regular, irregular
and hybrid verbs. In the literature, this distinction has not yet
been made because present tense conjugation has not been
essential, as only past tense and mostly non‐German languages
were studied. In the following, like in former studies, hybrid and
irregular verbs will be considered as the non-regular class. Hybrid
verbs are supposedly irregular verbs, only their stem vowel ([ei]
or [ie]) is not mutable to umlaut. The subsumption of irregular
and hybrid verbs is a simplification. The renaming of the irregular
class to non-regular should foster awareness for the simplification.
However, it is necessary to take up again with the finer grained
distinction in regular, hybrid and irregular verbs later on.
The next two chapters address issues concerning the cognitive

mechanisms of mental lexicon and mental grammar, i.e. the
dissociation of regular and non‐regular verbs. Chapter 2 provides
empirical evidence for the distinction and discusses models
attempting to account for these findings.
2 Approaches to regular and non-
regular inflection
2.1 Articulation latencies of regular and non-

regular verbs
This section presents a summary of previous studies

demonstrating that two types of knowledge entail two types of
cognitive processes. The first study using a speeded production
task on adult native speakers was presented by Prasada et al.
(1990, as cited in Seidenberg, 1992 and Pinker & Prince, 1994)9. In
the past‐tense naming task, participants generated past tense
forms of English verbs aloud and articulation latencies were
measured. Regular verbs were produced faster than non‐regular
ones, indicating a dissociation of cognitive processes. All
experimental items were visually presented as base forms. One
point of criticism against this outcome was put forward by Pinker
& Prince (1994) and targets the ratio of regulars and non‐regulars,
which was not balanced.
Similar results are reported by Seidenberg & Bruck (1990, as
cited in Seidenberg, 1992)10 who examined regular and non‐
regular verbs in a production task. Participants were randomly
presented with frequency‐matched regular and non‐regular
present tense verb stems, i.e. forms that are, in English, identical to
present tense inflected forms except in the third person singular.
9 Prasada et al. (1990) is a conference paper that, unfortunately, is not available

anymore.
10 The materials from the original conference presentation (Seidenberg & Bruck,
1990) proved unobtainable.
19
20 CHAPTER 2: APPROACHES TO REGULAR AND NON‐REGULAR INFLECTION
Participants first read the base form aloud and then generated the
corresponding past tense form. They were considerably faster
producing regular verbs than non‐regular verbs. Even after
subtracting the base‐form reading time from the past‐tense
generation time, it took participants significantly longer to
produce non‐regular verbs.11
Clahsen et al. (2004) modified the design of Prasada et al.
(1990) to study frequency effects on morphology in children. Of
interest here are the results from the adult control group: The
adults’ production latencies revealed a main effect for verb type –
a difference between regularly and non‐regularly formed
participles. As in previous studies, participles of regular verbs
were produced faster than non‐regular ones. Participants were
asked to inflect auditory stimuli (verb stems) into participles in
sentential context (e.g. Der Gitarrist hat die Gitarre *nehm-*
genommen. [The guitarist has *take* taken the guitar]). As a
qualification, it must be mentioned that the effect did not surface
in the item analysis. Though the phonological similarity (even
identity) of all item onsets (ge-) constitutes an advantage for
accuracy of measurement, a preparation effect could have
inhibited full development of the verb type difference. Still, the
latency difference was only about 20 ms – less than half of what
Prasada et al. (1990) and Seidenberg and Bruck (1990) had
observed. The main findings of the study indicated that the
mental mechanisms and representations for processing
morphologically complex words are the same in children and
adults. It was suggested that the production of non‐regular
participles may be slowed down by the retrieval of stored
participle forms from the mental lexicon.
In sum, all these studies employ similar experimental
techniques, and all report significantly longer reaction times for
non‐regular verbs. This result is unexpected assuming that non‐
regular verbs are stored as full forms in the mental lexicon. Lexical
11 Seidenberg & Bruck (1990) interpret the results as support for connectionist
models (see 2.4.1), although the pattern is no less compatible with Prasada et al.’s
(1990) study strengthening Dual Route processing (see 2.2).
2.2 DUAL ROUTE MODELS OF LANGUAGE PRODUCTION 21
look up, or the retrieval of full forms from memory, is supposed to

be faster than generating word forms by rule. In the following
section 2.2, the quality of psycholinguistic models on
representation and processing of verbs is measured by their
consistency with these empirical findings.
2.2 Dual Route models of language production
In research on monolingual speech processing, the models that

incorporate the idea of mental lexicon and mental grammar are
Dual Route Models (DRM). They have formed the basis for several
theories on verb production. The characteristic property of Dual
Route Models is the existence of two independent routes working
in parallel to attain a single outcome12. In the last decades, a
number of Dual Route Models have been proposed for both
production (e.g. the Words and Rules Theory; Pinker, 1999) and
comprehension (e.g. the Augmented Addressed Model (AAM);
Caramazza, Laudanna & Romani, 1988) of morphologically simple
and complex words. As the present discussion will be limited to
language production, the following review will focus on the well‐
known Words and Rules Theory by Pinker (1999).
The background of the Dual Route Models of verb production
is generative grammar. Early generative phonology (Chomsky &
Halle, 1968) tried to explain every linguistic expression through
rule‐based computation of symbols, except for totally intractable
forms like suppletives (e.g. (I) am). Dual Route Models of verb
production follow the traditional distinction between grammar
and lexicon: They combine the rule‐based single mechanism
model (e.g. Halle & Mohanan, 1985) and the associative single
12 Some Dual Route models do not broach the issues of morphology but other
cognitive processes like reading (see Coltheart, Curtis, Atkins & Haller, 1993). In
these models, learnt, familiar words which have a lexical entry are processed via
the lexical route. The lexical entry contains the phonological structure of these
known words. A non‐lexical route applies for reading of new or non‐words. Any
pronounceable letter string can be converted into sounds by using grapheme‐
phoneme conversion. Reading is a race between both routes.
mechanism model (e.g. McClelland & Rumelhart, 1986). Hence,

the crucial assumption in Dual Route Models is that there are two
routes reflecting different linguistic processes. Foremost, the
authors distinguish regular and non‐regular processing. Regular
forms are composed of stem and suffix by a rule‐governed
process. Non‐regular forms, however, are not derived by rule;
rather, they are stored in an associative memory system as ready
made lexical units.
The structural differences between regular and non‐regular
inflectional patterns reflect the very general characteristic of the
human language faculty: the division into lexicon and grammar
(Pinker, 1999; Pinker & Ullman, 2002; Clahsen, 1999). Irregular
verbs have signatures of lexical memory and regular verbs have
signatures of computational processes (Pinker & Ullman, 2002).
Phenomena like inflecting neologisms or overgeneralisations etc.
are empirical tests for the division of human language faculty and
will be discussed in detail as along with their neuropsychological
underpinnings in section 2.3.2. Several terms exist in the literature
for the regular/non‐regular13 dichotomy: Clahsen (1999) refers to
the distinction as lexically based inflection vs. inflection based on
combinatorial rules. Zwitserlood and colleagues (2002) use full
listing vs. decomposition approaches. Ullman (2001b) distinguishes
between declarative and procedural knowledge. His work is in fact an
implementation of the Dual Route Model in a broader theory of
language and memory. These models are grounded in
psycholinguistic data indicating two different cognitive
processing mechanisms.
2.3 The Words and Rules Theory
In Words and Rules (WR; Pinker, 1999), Pinker describes the

representation and processing of regular and non‐regular
13 The terms “regular” and “irregular inflection” (Pinker, 1991) are adapted to the
concept of regularity used in this work. “Irregular” denotes a subgroup of non‐
regular inflection.
2.3 THE WORDS AND RULES THEORY 23
morphology in English. He had already devised a hybrid model

for this in the early 80s, asserting that two different mechanisms
apply. In his description, he employs elements of associative
models as well as elements of strictly rule‐based accounts. For the
first of the two routes, the lexical route, Pinker claims that a
lexicon exists and that fixed words and forms are retrieved
without any opportunity to combine them in an unprecedented
way. This route is strictly limited to lexical look‐up of ready‐made
entries. The other route operates by applying a rule that
concatenates, for example, verb stems and affixes and will,
henceforth, be referred to as the rule‐based route. Both routes
activate lexical entries, but regular stems require the combination
of more than one element to form a complex word, and this
combination is achieved by means of grammar. The scope of the
rules is restricted to the inner construction of words (morphology)
and does not account for the serial order of words within
sentences14.
According to WR, humans do not possess a discrete rule for the
past tense generation. All the WR theory postulates is “a general
operation for merging or unifying constituents” (Pinker & Ullman,
2002). The “operation” is so “general” that it can also apply for
plural inflection of nouns where, like for past tense generation, an
affix is added to a stem. Actually, pluralisation and past tense
generation differ in input and output, but to concatenate
constituents the same rule applies. The great advantage of the
general merging operation becomes evident if the lexical route
breaks down. Then, the rule‐based route comes to the rescue.
Humans are able to apply the rule whenever memory fails, be it
due to aphasia or imperfect acquisition of the lexicon, and
unimpaired adults easily generalise rule‐driven processes to form
novel words.
In spite of the strict distinction of the lexical and rule‐based
route, the WR theory allows frequently used regular words to be
14 The output of both the lexical and the morphological route is input for syntactic
rules but it is beyond the scope of WR and this thesis.
stored in the lexicon, too, and to treat and access them as full
forms just like non‐regular words.
There can be no doubt that WR has been an influential work in
the field of human language processing; yet, it only makes explicit
predictions with respect to regular and non‐regular past tense
production. Obviously, the restriction to past tense does not allow
drawing generalisations to other tenses. German, to mention just
one inflectional language, possesses a much richer morphology
and regularity plays a role in suffixation of number/person
marking in present tense as well. Pinker’s discussion does not
extend to present tense. Therefore, the relation between present
and past tense still remains an underinvestigated issue.
2.3.1 The blocking mechanism
So far, a model containing a lexical route and a rule‐based route

would predict that non‐regular verbs are produced in a shorter
time than regulars because look‐up is faster than generating
words by rules. But empirical findings paint just the opposite
picture. Pinker & Prince (1994) and Pinker (1999) respond to the
observation that it takes longer to produce non‐regular verbs than
regular verbs (Prasada et al., 1990; Seidenberg, 1992) by
introducing a mechanism of blocking: A non‐regular form (e.g.
came) prohibits application of the rule in order to avoid
ungrammatical forms (e.g. comed) (also Pinker, 1999:130). Blocking
the regular route is costly and time‐consuming. The production of
irregular forms takes longer because blocking demands additional
cognitive resources.
The term blocking was coined by Aronoff (1976), originally
denoting a semantic principle that serves to avoid synonymy. If a
society has a viable expression for a particular concept, there is no
need to invent a new expression for the exact same concept. For
example, it is perfectly fine to say this morning, but *this night is
not, because there exists the expression tonight which refers to the
intended concept. In this context, blocking is conceived both as a
synchronic theoretical principle and as a diachronic tendency.
Likewise, availability of a viable non‐regular verb form which
predates a regular verb form expressing the same concept will

block regularisation (e.g. ran blocks runned). Furthermore,
blocking is not restricted to inflection: Aronoff discusses the term
in the context of derivation as well, for cases where the same base
is involved (brevity blocks brevitude). The role of frequency in
blocking is crucial: the more frequent item blocks the less frequent
one (Aronoff, 2000).
Psycholinguistics essentially uses a variant of the blocking
principle. Analogously, the blocking principle in morphological
processing dictates that a non‐regular form listed in the mental
lexicon blocks the application of the corresponding general rule,
i.e. the generation of a new word. The non‐regular form broke
prevents adults to add the regular suffix –ed to break saying
breaked. In the WR framework, the regular and non‐regular routes
are initiated by default for all verbs and operate in parallel (parallel
race model). If the lexical route yields a successful result prior to the
rule‐based route – look‐up is a faster process than composition at
any rate –, retrieval of the pre‐stored inflected form blocks (“pre‐
empts”) the concatenating rule‐based route. However, because of
higher cognitive processing load, blocking the rule‐based route is
time‐consuming. Therefore, the production of non‐regular forms
takes longer. Should the lexical route not lead to a suitable form,
the rule‐based route wins and applies by default.
Sahin et al. (2006) claim that they have found a neural correlate
for blocking in an fMRI study. They made English native speakers
inflect or read words sub‐vocally. The experiment had a 2 × 2 × 3
factorial design, manipulating Grammatical Category
(Noun/Verb), Regularity (Regular/Non‐regular) and Task (Overt‐
Inflect/Zero‐Inflect/Read). The participants’ responses were cued
by sentence fragments, like Yesterday they __ calling for past tense
and, hence, overt inflection as opposed to zero inflection with a
null morpheme. Utterances in the reading condition were elicited
by Read word: ___.
For the purposes of the present work, comparisons subtracting
regular from non‐regular overt inflection of nouns and especially
verbs are of primary interest. The fMRI‐activation patterns of this
particular comparison were not conforming to the well‐known
anterior (non‐regular) / posterior (regular) dissociation (see next

section 2.3.2). Although the experiment failed to exhibit the typical
dissociation, non‐regular processing increased activation
markedly in the supplementary motor area (SMA) in both
hemispheres. Sahin et al. (2006) report that this region is
associated with conflict resolution or inhibition of habitual
processes. Therefore, they attribute the region’s activity to
suppression of a morphological rule, which means the area is
activated either for blocking processes or for competition between
similar non‐regular forms. The latter alternative is more likely,
and it supports the assumption that this correlate for blocking is
not as pronounced for nouns as it is for verbs because the number
of competitors is lower.
Recordings of participants’ fixations during inter‐trial intervals
were intended to serve as a baseline, but in their data analysis the
authors opted for subtraction of conditions from each other, such
that, for example, overt inflection (plural or past tense) was
contrasted with zero inflection. Sahin and colleagues (2006) argue
that the difference in activation between both conditions reflects
the retrieval of phonological content of the suffix, plus
concatenation of stem and suffix, plus phonological adjustments.
Still, once the study is more carefully examined, it becomes
disputable whether the correlate really reflects blocking.
Methodological objections concerning subtraction include that any
analysis by subtraction faces the challenge of establishing valid
minimal pairs baseline vs. baseline + task. It is very difficult to meet
this challenge in practice, and even then, the interpretation of the
results remains full of uncertainties. In Sahin et al. (2006), the
authors did not subtract a condition from a baseline but two
conditions from each other. Obviously, different conditions
employ non‐identical tasks and to date we have only vague ideas
about the exact processing mechanisms behind each task. It is well
possible that overt inflection and zero inflection differ in more
than what we associate with suffixation (as explained above).
Therefore, it remains an open question what really is reflected by

the activation pattern in Sahin et al. (2006)15.
Separate comparisons of zero inflection and overt inflection for
nouns and verbs would have been of highest importance to the
current study. As Sahin et al. (2006) studied English, a comparison
between zero inflection and overt inflection of verbs would have
been equivalent to a present and past tense contrast. This
interaction between word class and inflection is not resolved, or at
least not reported, even though the authors themselves reported
differences in the processing of nouns and verbs.
The reader should keep these objections in mind. Arguments
against blocking will be given in section 2.3.4.
2.3.2 Psycholinguistic evidence for a regular/non-regular

dissociation of verbs
A large body of research supports the lexicon vs. grammar

dissociation of regular and non‐regular verbs, including findings
drawn from:
(a) past‐tense acquisition in children (Marcus, Pinker, Ullman,

Hollander, Rosen & Xu, 1992)
(b) normal adult processing (Stanners, Neiser, Hernon & Hall,
1979; Alegre & Gordon, 1999; Prasada et al., 1990)
(c) breakdown of speech production in aphasic patients
(Marslen‐Wilson & Tyler, 1997; Ullman, Corkin, Coppola
& Hickok, 1997) and
15 During the last years, a debate arose whether neuroimaging studies contribute
to the development of theories and whether they enhance our understanding of
cognitive process at all. Neuroimaging studies use the blood‐oxygen‐level
dependent (BOLD) signal as dependent variable but it is not entirely clear what an
increase in the BOLD signal reflects and means (Logothetis, 2008). Coltheart (2005)
and Page (2006) doubt any substantial contribution of neuroimaging techniques
because the data provide no evidence to differentiate between possible models.
Cognitive models cannot be tested using neuroimaging data because the functional
role of parts of the brain is as yet not sufficiently understood (Page, 2006). Pure
localisation of cognitive processes is not what cognitive scientists are interested in.
They want to know how the brain works.
(d) neuroimaging and electrophysiology (Jaeger, Lockwood,

Kemmerer, van Valin, Murphy & Khalak, 1996; Beretta,
Campbell, Carr, Huang, Schmitt, Christianson & Cao,
2003; Newman, Ullman, Pancheva, Waligura & Neville,
2007).
2.3.2.1 Past-tense acquisition
The acquisition of non‐regular words follows a U‐shaped learning

curve that plots the percentage of correctly produced past tense
non‐regular words: correct production is followed by a phase of
incorrectly regularised non‐regulars. After a while, children make
fewer mistakes, until they eventually produce nearly all regulars
and non‐regulars correctly (Marcus et al., 1992).
The proportion of those over‐regularisations is relatively small,
but the existence of the phenomenon confirms the assumption of a
symbolic rule‐based mechanism: children extract the rule for past
tense formation from their language input (“append -ed”) and
tend to over‐apply it to other forms. This happens only when the
children have not yet acquired the non‐regular form that could
block the application of the rule. This period of over‐
regularisation requires recognition and awareness of rules in the
language. Therefore, incorrect realisation is not a step backwards
after children have already produced correct past tense forms:
initially, they had stored all forms in the lexicon. Later, with
temporarily increasing error rates, they apply rules.
2.3.2.2 Normal adult processing
Priming
Many behavioural studies in the large body of literature have

employed priming tasks. If a target word is preceded by a prime,
reaction times to the target stimulus are demonstrably affected by
the type of relation that holds between prime and target. These
relations are then manipulated experimentally. Influential
priming‐based work related to the issue at hand has been
undertaken by Stanners et al. (1979). In a lexical decision task,

regularly inflected words primed (facilitated) their stem16 to the
same extent as stems primed themselves (full‐priming). The
priming effect on words such as walk was as large when it was
preceded by walked as when it was preceded by walk. The prime
walked and the target word walk are morphologically related. With
such pairs, reaction times for lexical decision about the target are
faster than for pairs where prime and target are not related.
Overall, studies have shown that facilitatory priming effects for
regular verbs are consistent (e.g. Stanners et al., 1979; Sonnenstuhl,
Eisenbeiss & Clahsen, 1999). According to Dual Route theories,
they are decomposable and are in fact decomposed during
processing (walked = walk + ed). The prime walked activates the
lexical entry for the stem walk, so the target is activated and
retrieved more easily. More precisely, for regular verbs, the stem
primes itself (Pinker, 1997). However, non‐regular verbs primed
by their past tense forms did not produce an equivalent degree of
facilitation.
Inflected regular past tense forms are orthographically and
phonologically more similar to their stem than non‐regular past
tense forms are (see, for example, walked ‐ walk versus taught ‐
teach) and it might be these different form properties that account
for full priming of regular past tense forms. However, facilitation
cannot be explained by effects of orthographic priming (Stanners
et al., 1979; Rastle, Davis, Marslen‐Wilson & Tyler, 2000).
In a control experiment, Stanners et al. (1979) investigated the
effects of orthographic similarity of irregular past tense forms on
morphological priming. In one of their experiments (Experiment
2), prime and target differed in the amount of similarity: prime
and target differed either by one letter (the similar condition; e.g.,
hang–hung) or by more than one letter (the dissimilar condition;
e.g., shake–shook). As the priming effect persisted, it can be
concluded that the difference in reaction time does not stem from
orthographic overlap – at least not entirely. Under the Dual Route
account, regular verbs have to be assumed to have only one lexical
16 In English, the stem of verbs is identical to the infinitive.

entry. By contrast, non‐regular verbs showed only reduced

priming effects. They seem to be not morphologically related to
their base form.
However, results are inconsistent concerning the question of
priming effects for non‐regular verbs; for these, researchers found
either reduced/partial (Stanners et al., 1979; Clahsen, 1999) or no
priming (Kempley & Morton, 1982). The inconsistencies might be
due to partial regularities of non‐regular verbs. In general, non‐
regular verbs do not exhibit full priming because they are not
stored in a decomposed fashion, but have a special past tense
form. Hence, the mental representations of the two verb types
must be distinct.
Frequency effects
Generally speaking, reactions to more frequently used words are

easier and faster performed (Whaley, 1978, in lexical decision
tasks; Forster & Chambers, 1973, in naming tasks). Still, such
frequency effects are only observed with processing of non‐
regular word forms. It is known from Clahsen (1996) that word‐
form frequency matters, such that non‐regular processing is faster
for high‐frequent words than for low‐frequent words. In a visual
lexical decision task, the stem frequency was held constant.
However, word form frequency effects were observed for non‐
regular German participles (e.g. ge-sung-en [sung]) but not for
regular ones (e.g. ge‐lach‐t [laughed]). The frequency effect for
non‐regular participles has been interpreted as evidence for
distinct allomorphic stems of non‐regular verbs. Regular
participles did not exhibit such a frequency effect in Clahsen’s
(1996) study and thus are assumed to be represented as single
stem from which word forms are computed by rule.
Meanwhile, multiple studies support a reverse (or anti‐)
frequency effect for visual processing of regular words (Alegre &
Gordon, 1999; Clahsen et al., 2004), where LD‐responses to high‐
frequency regulars are slower than to low‐frequent ones. Beck
(1997) had previously reported this finding, but had presumed it
to be an artefact resulting from the experimental design. Clahsen
et al. (2004) explain the results with the Words and Rules Theory
by Pinker (1999): Because of their widespread use, high‐frequent
regular word forms might be stored in the mental lexicon as well.
For reasons of economy, they are not computed with every lexical
access. Thus, the system avoids the generation of identical items.
High‐frequent regular verb forms share their representational
status with non‐regular forms in the Dual Route Mechanism.
Stored forms block the application of the rule and slow down
reaction times.
Similar to visual LD experiments, experiments of speeded
production of inflected word forms revealed dissociating
frequency effects. In a past tense naming task by Prasada et al.
(1990, cited after Seidenberg, 1992, and Pinker & Prince, 1994), the
frequency of the base form influenced both regular and non‐
regular past tense generation, while the frequency of the past
tense form affected generation of non‐regulars only. The
frequency effect with longer latencies for low‐frequent words for
non‐regulars (but not for regulars) has since been replicated for
example by Seidenberg & Bruck (1990, as cited in Seidenberg,
1992). With respect to the representation of regular and non‐
regular verbs, similar conclusions are drawn from production as
from reception. Reaction times favor a Dual Route representation.
2.3.2.3 Breakdown of speech production in aphasic patients
Agrammatism is a common symptom in the course of brain

damage to Broca’s area. Agrammatic speech is difficult to initiate
and lacks fluency. While comprehension in agrammatic patients is
generally preserved17, Broca’s aphasia affects production of words
and sentences. Agrammatic Broca’s aphasics exhibit a deviant
pattern of verb production. The first thing to notice in their
spontaneous speech is the absence of verbs and second, that the
verbs produced lack inflection.
In a cross‐modal priming task (Marslen‐Wilson & Tyler, 1997),
English native patients with agrammatism as a result of Broca’s
17 although there are also aphasics with receptive agrammatism

aphasia were selectively impaired in their performance on English

past‐tense lexical decision. Some showed partial priming for non‐
regular verbs while priming of regular verbs was absent; one
patient displayed the reverse priming pattern. The authors of the
study took this as evidence for a regular/non‐regular double
dissociation indicating different neural structures involved in verb
processing. Complementary data revealed that anterior lesions are
often accompanied by impairments of regular inflection, posterior
lesions by impairments of non‐regular inflection (Ullman et al.,
1997).18
In a study of patient AW, Miozzo (2003) attributed the
dissociation to a more general principle. AW’s deficit did not
affect regular verbs (walked) or nouns (gloves), but her performance
with respect to non‐regular forms (found as well as children).
Morphological rules seemed to be spared because AW’s
impairments were not restricted to one single word class, but
rather affected particular types of inflection. Dual Route
Mechanisms can be generalised over word classes, though
properties of word classes vary in several respects. In general,
processing of regular inflection may remain normal, while
processing of non‐regular inflection may be selectively damaged
and vice versa.
Researchers have different opinions as to whether the existing
neuropsychological and neuroanatomic data are convincing. In a
meta‐analysis of 75 individuals with Broca’s aphasia by Faroqi‐
Shah (2007), performance of regular/non‐regular verb production
is characterised by heavy inter‐ and intra‐individual differences.
The meta‐analysis shows that the dissociation between regular
and non‐regular verbs is not robust and lesion patterns are not
consistent. The great variation in the data is rooted in the variety
of methods, statistical analyses, concepts, operational definitions,
and most importantly, handling of control subjects.19 Furthermore,
18 past tense generation task: Patients were presented with a sentence like “Every
day I dig a hole. Just like every day, yesterday I ____ a hole” and asked to complete
the sentence by providing the past tense form of the verb (here: “dug”).
19 Ceiling effects for control subjects present statistical problems.
two objections to the meta‐analysis can be made: First, Faroqi‐

Shah (2007) analysed more than 100 data sets without
disentangling the locus and type of impairment carefully enough.
Indeed, difficulties in retrieving the stem of regular verbs is also a
memory‐dependent capacity like the retrieval of non‐regular
verbs. In this regard, it actually appears reasonable not to expect
dissociations between regular and non‐regular inflection. Why
then should the lesion sites of regular and non‐regular verbs
dissociate? Second, patients’ improvements over time are often
disregarded. Simmons‐Mackie & Damico (1997) have shown that
aphasic patients often develop compensation strategies for
lesioned parts of the processing system, so data from their
processing performance may not be as convincing as it seems at
first glance.
Single brain areas indexed by lesion studies are necessary, but
at the same time not sufficient for particular operations. In any
case, this section shows that regular and non‐regular verbs can be
selectively impaired. Thus, there seems to be a double dissociation
between regular and non‐regular past tense forms in these
patients, which indicates that the processes underlying regular
and non‐regular inflection can be selectively impaired and
therefore belong to different cognitive systems (Marslen‐Wilson &
Tyler, 1997).
2.3.2.4 Neuroimaging and electrophysiology
If the production of regular and non‐regular verbs follows

different principles, it is plausible that the relevant processes take
place in different brain areas (Marslen‐Wilson & Tyler, 1998).
Indeed, imaging studies show different patterns of activation for
processing of regular and non‐regular forms. Basically, non‐
regular verbs activate larger brain regions than regular verbs
(Jaeger et al., 1996; Beretta et al., 2003, among others); in
particular, activation occurs in distinct, partially overlapping
areas: while the processing of regular forms is more left‐
lateralised, non‐regular processing exhibits a tendency towards
bilateral activation.
Across studies focusing on the cortical regions for processing

regular and non‐regular verbs, the two major regions involved in
differential activation are the temporal and frontal areas (see a
review in Beretta et al., 2003). More precisely, it is the left inferior
frontal cortex, Broca’s area, which shows activation for processing
and production of regular verbs. In contrast, abstracting over the
multitude of studies, posterior areas (left tempo‐parietal cortex)
subserve the processing and production of non‐regulars (Jaeger et
al., 1996). This strongly supports the Dual Route hypothesis and is
consistent with neuropsychological evidence (Ullman et al., 1997).
However, exact activation patterns do not match across
studies. There are no two studies with comparable results. The
numerous studies form a heterogeneous picture with respect to
specific areas associated with either type of inflection. It may well
be because methodologies varied from PET to fMRI and languages
from morphologically rich to poor. Further, experiments
presented stimuli either in a blocked or mixed design, or tested
only very few participants and, finally, experiments employed
various different tasks (silent naming, repetition, inflection) and
data analyses (comparison of conditions or subtracting out activity
from neutral condition).
Neuroanatomical correlates spread over a wide range, and
therefore are inconclusive. The advantage of the imaging
technique is its high spatial resolution, but due to its low temporal
resolution, the method cannot tap the time course of inflection
processing. In fact, reconsidering the aforementioned findings,
they are suited to arouse the suspicion of a paradox: greater
activation reflects and implies higher processing load. Despite
regular verbs being morphologically complex and non‐regular
verbs morphologically simple, the evidence suggests that it is
more demanding to produce or retrieve non‐regular verbs, even
though the processing of regulars must proceed in at least two
steps, that is, one more than what is necessary for non‐regulars.
Support for the frontal/temporal distinction also comes from
event‐related brain potential (ERP) studies recording electric
potentials from the scalp (cf. Newman et al., 2007, and Ullman,
2001a, for a review of several studies). Violations of regular
suffixation rules (omission or misapplication) elicited a Left

Anterior20 Negativity (LAN), and in some studies triggered the
P600 component (Lück, Hahne & Clahsen, 2006). Roughly, both
are associated with syntactic processing. In most studies, the LAN
reflects an early, purely syntactic process (Friederici, 1995). The
P600 component is a positive voltage deflection 600 ms after
stimulus onset. The amplitude increases in response to verb
agreement violations (Osterhout & Mobley, 1995) or phrase
structure violations (Hahne & Friederici, 1999). For morphological
violations, the same components manifest themselves as for
syntactic ones. Manipulations of non‐regular morphology elicited
an N400 in a classical bilateral fashion, distributed temporally
(Penke, Weyerts, Gross, Zander, Münte & Clahsen, 1997). Grossly
simplified, the N400 reflects lexical‐semantic integration and
processing. Hence, the ERP‐patterns of incorrectly inflected verbs
support the distinction between mental lexicon and mental
grammar of morphologically complex words.
2.3.3 Representation of regularity in the Words and Rules

Theory
So far, WR states that regularity does not have a representational

counterpart but emerges from the architecture of the Dual Route
Mechanism. Regular forms are produced if no non‐regular form
has been found after an exhaustive search (Pinker & Ullman,
2002). There is no need to postulate an explicit feature
representing the regularity status of words.
“From our brain’s point of view no verb is either regular or irregular

until it has been looked up in memory and discovered to have, or to
lack, a special past tense form” (Pinker, 1999:131)
20 The regions correspond to the imaging studies. There is mutual support

between ERP and imaging studies.
Verb production is governed by two distinct subsystems that

allow a binary choice and no deviations. This is the foundation for
the generalisation that regular and non‐regular verbs “should be
dissociable from virtually every point of view” (Pinker, 1991:532). The
independence of these two systems can be confirmed by the
observation that they may be selectively affected by language
disorders (Miozzo, 2003).
In 2002, Steven Pinker elucidates his ideas concerning the
nature of morphological features. He claims that non‐regular
verbs are stored with a past tense feature in their lexical entry,
which specifies non‐regular forms as non‐regular:
“Irregular forms are just words, acquired and stored like other words,
but with a grammatical feature like ‘past tense’ incorporated into their
lexical entries” (Pinker & Ullman, 2002:456).
However, he is not very explicit and specific about this point.

Usually, grammatical information is represented as abstract
generic nodes. Schriefers (1993) and Jescheniak and Levelt (1994)
addressed the same question exploring the representation of
grammatical gender. The idea of generic nodes is that each item
with a particular property is connected to the same node. In fact,
Pinker’s suggestion is restricted to irregular verbs.
Still vaguely, WR (see Pinker & Ullman, 2002) predicts that all
words are connected in such a way that, in case of impairment, all
regularly inflected words will show a deficit, i.e. nouns and
present tense verbs alike. This would imply a specific
superordinate feature that is stored in the mental lexicon and for
which every lexical entry must be specified. One possibility to
achieve this is the idea of generic nodes, such as for gender or
grammatical class. All words, i.e. their corresponding lemmas, are
supposed to be linked to that particular feature. In favour of a
generic regularity feature, patient AW – suffering from a left
temporal lesion (Miozzo, 2003) – responded less accurately to non‐
regular forms of both verbs (e.g. found) and nouns (e.g. children). In
her case, the deficit in producing non‐regular inflection clearly set
apart regularly from non‐regularly inflected words. Her
impairment affected past tense, past participles and plural

formation. All mechanisms associated with regular processing
were seemingly intact, since they yielded the same (appropriate)
kind of responses as controls.
Miozzo suggested AW’s deficit to be one of retrieving the word
phonology from the lexicon. AW did not pursue the strategy add –
ed to a verb to form its past tense because she also guessed non‐
regular verb forms by changing the vowel (see also Patterson,
Lambon Ralph, Hodges & McClelland, 2001). Hence, she knew
whether a verb takes a non‐regular form or not but non‐regular
phonological forms were inaccessible. This fact could lead one to
conclude that the information about a verbs’ regularity status is
stored separately from its phonological form. In reference to
Pinker (1991), Miozzo (2003) envisions a representation of a
regularity feature and ascribes it the important function “to block
the suffixation process and to trigger the retrieval of the irregular [non‐
regular, H.T.] form from stored phonology.” (Miozzo, 2003:124).
There has repeatedly been evidence that regular and non‐regular
inflection actually behave differently, but the mechanism behind
them has not been addressed carefully. It will be tested here
whether the notion of a regularity feature could help to clarify
empirical findings related to the regular/non‐regular distinction.
2.3.4 The Words-and-Rules-Theory’s difficulties
WR has two central limitations: first, it makes explicit predictions

only in relation to regular and non‐regular past tense production.
Hence, tense is a non‐considered factor. Second, the model is
limited, as it concentrates solely on English as the language of
interest. Research on the interaction of inflection (person/number)
and present tense would be illuminating in itself, and it would
also help to disentangle the confound resulting from the formal
overlap of person/number/tense inflectional endings in English
verb inflection.
One source of evidence for WR was a pronounced frequency
effect for non‐regular forms as opposed to an absent frequency
effect for regular verbs (Prasada et al., 1990). However, frequency
effects should be observable for regular inflection as well: the

stems of regularly inflected verbs also have to be retrieved from
the lexicon and underlie the same psycholinguistic mechanisms
for lexical access. Consequently, a valid position could be to
expect frequency effects within paradigms in non‐regular verbs
and between paradigms of regular verbs, as they employ the same
stem within a paradigm.
The following problems arise out of the limitations: first,
blocking is a promising attempt by Pinker & Prince (1994) and
Pinker (1999) to explain empirical data. However, according to the
WR blocking is kind of waiting of a non‐regular form for spell out:
a quite unintuitive and uneconomic mechanism. The blocking of
the rule‐based route can hardly be a general cognitive principle. It
is like abandoning the mental grammar and this would at the
same time put the creativity of the language at risk. An example
should demonstrate that blocking can be bypassed: humans are
able to modify idiosyncratic expressions on the level of both
morphology and syntax. For example, although (er) *geh-te [he go‐
ed] even (er) *ging-te [he went‐ed] are ungrammatical, they are
possible to produce and are even interpretable correctly. No rule‐
based route is blocked despite the existence of the non‐regular
form (er) ging [he went]. The neural correlate for blocking found
by Sahin et al. (2006) is not considered striking evidence due to the
methodological problems discussed above.
Secondly, the model is not precise about the representation of
regularity (see section 2.3.3): It stores regular verbs and non‐
regular verbs in a shared lexicon. Whenever stems are retrieved
from the lexicon, both routes are active and run in parallel. It is
not clear from the model how the appropriate route is triggered,
i.e. where the information comes from whether the rule‐based
route has to be blocked or not.
2.4 Connectionist accounts
During the 80s, some researchers dismissed symbols and rules

and put forward single‐route models and explanations.
2.4 CONNECTIONIST ACCOUNTS 39
Connectionist models are the most widely discussed single‐route

models (e.g. McClelland & Rumelhart, 1986; Plunkett &
Marchman, 1993). Others are the full‐listing theory by Butterworth
(1983) and Bybee (1988) or the Analogy proposal by Skousen
(1992). Representing other single‐route accounts, connectionist
models shall now be introduced in more detail. The original rule‐
less model is the Pattern Associator, a Parallel Distributed
Processing model by Rumelhart & McClelland (1986)21.
Connectionism denotes a school of thought within cognitive
science, in the field of artificial intelligence. It focuses on the
development of cognitive skills such as language acquisition.
Researchers study the self‐organising learning process the final
state of which is the object of empirical investigation.
Connectionist networks constrain learning algorithms and
simulate humanlike neural processes with the help of computers.
Advocates of connectionism assert that behind knowledge and
cognitive skills (such as language) stand associative networks
rather than symbols, rules and combinatorial operations. They
follow the basic principle of describing mental processes through
interconnected networks of simple units. Multiple sources of
knowledge and interacting mechanisms are assumed. Networks
can structurally resemble neurons and synapses, or words with
their manifold relations. The system provides an environment and
resources from which (grammatical and lexical) knowledge can
emerge.
Single‐route models have in common that they treat both
regular and non‐regular verbs in the same way. Words are stored
associatively in one single memory system, none are decomposed.
Even regular past tense verbs are represented as full forms,
without using explicit (i.e. symbolic) rules.22 Regular stems are
21 Single‐route models must not be put on a level and associated with

connectionism. Nothing speaks against a connectionist model that explicitly
consists of two routes or that develops two routes automatically during training
and learning.
22 However, connectionist models can indeed employ explicit rules. Rodriguez et
al. (1999) demonstrate a neural net which has learnt how to count. Its acquired
knowledge can be described and represented by rules. However, it is often not
mapped directly onto their past tense form through generalisation

procedures. Generalisations of mapping are mainly driven by the
frequency of co‐occurrences and similarity between items.
Connectionist models challenged the classical theory of mind,
according to which cognition relies on symbol manipulation by
formal rule application. Critical remarks followed by Levelt
(1991). He compared the two approaches to the theory of mind at
hand: the symbolic and connectionist one. Connectionist models
are finite state machines with a predefined finite number of states
(see Levelt, 1991). On the downside, current connectionist models
cannot handle the symbolic aspects of language. They do not
represent adequately how a finite number of elements of a
particular language can be used infinitely. Therefore, the
connectionist account has its limits e.g. in recursion. This makes
connectionist models intolerant towards expansion of knowledge
and can only represent a well defined section of human cognition.
2.4.1 The Pattern Associator
The most popular associative single‐route model is the Pattern

Associator conceived by Rumelhart & McClelland (1986) for
English past tense inflection. A pattern associator learns
associations between input patterns and output patterns, e.g. in
case of the Pattern Associator between past tense forms and
infinitives. Most appealing, it can generate what it learns about
one pattern to similar input patterns.
The network consists of nodes and connections. Nodes process
incoming activation and, depending on their own activation,
nodes influence connected nodes. They do not pass symbolic
information but deliver numeric activation values which are a
result of computations.
A neural network consists of many simple processing units
each of which is connected to many other units. Basically, the
Pattern Associator is a two‐layer network: it consists of input
easy to reveal the rules because they are intertwined with weights and activation
values in the network.
nodes and output nodes. Each node in a layer is connected to all of

the other nodes in its layer, and all of the nodes in each adjacent
layer, e.g. past tense forms are connected to infinitives. Each unit
has a numerical activation level (analogous to the firing rate of
real neurons). Input units in a network receive information, while
output units send information. The activation of an output unit is
determined by the product of the input activation and the specific
weight associated with that connection. The connections are
uniform and not labelled. However, connections differ in their
strengths. The relations between nodes emerge through statistical
computations: Training, i.e. increasing input, means adjusting and
modifying connection weights and finally learning. With every
usage, connections become stronger, while unused ones remain
weak. The strength of connections arises and increases also by
similarities (e.g. semantic, orthographic or phonological) to
neighbouring nodes. The Pattern Associator can be trained to
respond with a certain output to a presented input. Adjusting
connection weights changes input/output behaviour.
The Pattern Associator strives after a direct mapping from the
phonetic representation of the stem to the phonetic representation
of the past tense form. It works as follows: first, an input‐encoding
network takes a phonological representation of a stem and
converts it into Wickelfeature format. To model the linear order of
phonemes within verb stems, for example, Rumelhart &
McClelland (1986) introduced so called Wickelphones, which are
trigrams of phonemes. The verb walk is for example represented
by the four Wickelphones #wa, wal, alk, lk#. This way, the order of
phonemes in a word is coded unambiguously. The Wickelphones
themselves consist of Wickelfeatures which combine the
distinctive phonetic features of the Wickelphones reducing the
number of nodes. Then, verb stems are paired with the
Wickelfeature representation of past tense in the Perceptron
Pattern Associator. An output‐decoding network decodes the
structure back into phonological representations. By the
Wickelfeature format, the model uses distributed representations.
Words are represented by several nodes and each node is
involved in the representation of several words.
The phonological relationship between present and past tense

of non‐regular forms is arbitrary. Although rhyming schemes of
phonologically related non‐regular verbs do exist by similarity
(e.g. sink - sank, drink - drank), these relationships between past and
present tense stem are not reliably predictive. There are many
exceptions and counterexamples (e.g. think - thought, link - linked,
not thank or lank). However, the existence of these conjugation
classes is in perfect accordance with connectionist views of the
Pattern Associator.23 Even for processing non‐words, the Pattern
Associator has learnt what the most suitable form would be. It
works with abstract strings rather than words.24
One of the goals of Rumelhart & McClelland (1986) was to
model the acquisition of the English past tense. The Pattern
Associator indeed captures lots of the intricacies of this process
without rules but only by retrieval of full‐listed forms. The output
converges well with acquisition data of children. Errors made by
the model do not only resemble errors made by children, they
even follow the U‐shaped learning curve of non‐regular verbs.
The outcome of the experiments introduced in 2.1 above was
that regular verbs are responded to faster than non‐regular verbs.
The explanation provided within the connectionist framework for
the naming latency difference between regular and non‐regular
verb production does not only depend on properties of the words
itself, e.g. its frequency, but also on its neighbourhood, i.e. the
similarity between patterns of spelling of connected words.
Greater phonological similarity of a novel or a real verb stem to
other real non‐regular verb stems increases likelihood and
acceptability of non‐regular past tense forms. In contrast,
phonological similarity to regular verbs does not have any effect
(Bybee & Moder, 1983). Furthermore, regular and non‐regular
verbs’ neighbourhood differ in their consistency, i.e. non‐regular
verbs are surrounded by more exceptions (“enemies” in
23 Not all connectionist accounts are based on phonological information. Instead,

the relationship between present and past tense non‐regulars tends to rely more
heavily on semantic information (e.g. Joanisse & Seidenberg, 1999).
24 Therefore, it is not sensitive to word frequency.
Seidenberg, 1992) than regular verbs are and therefore yield

longer reaction times. Why does the Pattern Associator perform
like it does? Weights of connections between nodes encode facts
about frequency and consistency of correspondences. The more
frequent a verb is, the stronger the weights become. The more
similarities two words have, the stronger their connection.
Connections in the network are weighted and are a product of
learning algorithms. As regular verbs have many friends
(moderately deviating neighbours) and less enemies (strongly
deviating neighbours) connections between regular verbs are
stronger and therefore the nodes have more activation. These
words can be spelled out very easily. Non‐regular words have
only few friends and many enemies. The neighbourhood of non‐
regular verbs is diverse and inconsistent. Thus, because of less
pre‐activation, non‐regular verbs yield lower reaction times (cf.
Seidenberg, 1992).
However, none of the current connectionist models is entirely
psychologically realistic (cf. Ling & Marinov, 1993, for a critical
review). In the Pattern Associator, homophones cannot be
processed accurately because the model relies on phonology. How
does the system know which form to select or produce? Let us
consider the English example to lie. Its past tense can be either
regular (lied, lied) meaning “make false statements” or non‐regular
(lay, lain) to express “place oneself in a flat position”. MacWhinney
& Leinbach (1991) made an attempt to distinguish past tense
forms of homophones like ring and wring. The model used a new
architecture and learning algorithm as well as a larger input
corpus, a fuller paradigm, and a new phonological representation.
Learning the past tense was more realistic. MacWhinney and
Leinbach (1991) solved the homophony problem by including
semantic features that allowed the system to differentiate between
homophones.
2.4.2 Strengths and weaknesses of connectionist models
Connectionist accounts have often been criticised, and were

continuously optimised in reaction to criticisms (cf. Pinker &
Prince (1988) and Marcus et al. (1992), evaluating McClelland &
Rumelhart, 1986). Their strong point is trying to explain how the
brain processes information. Neural networks represent the brain
and cognitive processing reliably and incorporate the flexibility
demanded by the complexity of the real world. Learning
algorithms are robust, i.e. they can tolerate minimal deviations
from materials on which they were trained. Connectionists argue
that neural networks can handle phenomena naturally and they
provide ample simulations for cognitive tasks like object
recognition, action planning, and coordinated motion. Neural
networks are a good basis for operations that require parallel
processing. They retrieve information quickly and process it
effectively.25
Irrespective of the achievements of current connectionist
models, connectionists are continuously faced with substantial
problems. As the neural networks function on the basis of
resemblance and similarity, it is difficult for them to abstract from
learnt criteria the model is based on. Thus, concepts are handled
with only limited flexibility: according to a neural net, a tiger
could be defined as a black‐and‐orange feline. Albino tigers,
however, would not be recognised as tigers under such a
definition, as they do not meet the criteria for “tiger”.
Furthermore, the models simplify biological diversity like
physiological differences between different types of neurons, or
between neurochemical processes. The models are not suitable for
yet higher cognitive processes like emotions and reasoning thus
far (cf. Levelt, 1991).
25 However, parallelism could be captured by symbols‐and‐rules models as well.

2.5 SUMMARY 45
2.5 Summary
From the above review of empirical data and models of past tense
generation, it becomes evident that the past tense debate has
reached stalemate. Dual Route Models are built on the assumption
that regular and non‐regular verbs behave differently. Indeed, the
studies reviewed imply that two qualitatively different systems
are involved in verb processing. Likewise, connectionists do not
claim that verbs must necessarily behave alike just because they
are stored in a single network. Münte and colleagues (Münte,
Rodriguez‐Fornells & Kutas, 1999) pointed out that also the
neuroscientific fMRI and ERP data support neither a single nor a
Dual Route Model. They believe the issue to be more complex
than “one or two computations”. Both classes of models leave an
incomplete picture.
The emphasis in this work is not on contributing to that
controversy by finding empirical evidence for or against either single
or Dual Route Models. Instead, the important conclusion to be drawn
from criticisms against the blocking mechanism is that German
provides a test case for the blocking mechanism if one is to consider
not only past tense, but also present tense. In German present tense,
(er) beiß-t [he bite‐s] versus (er) lach-t [he laugh‐s] are both built
regularly. However, beißen [to bite] is non‐regular in the past tense
and lachen [to laugh] is regular in the past tense. WR claims that
individual forms are regular or non‐regular and therefore would not
predict the classic regular‐non‐regular distinction in present tense
where both forms can be inflected regularly.
This thesis deals with these problems of WR and tries to
intervene through looking at German and the present tense.
Previous studies did not consider whether regularity is represented
as a property of individual forms (implicit assumption of DRM) or
of whole inflectional paradigms. The following experiments
investigated hybrid verbs to test these two hypotheses. The
representation of regularity is explored in Experiments 1‐4 and the
blocking mechanism is tested in Experiment 5. This raises a more
complete picture of German verbal inflection and sheds light on the
mental lexicon and grammar more broadly.
3 Psycholinguistic models of
language production
This chapter covers aspects of speech production that are relevant
to the discussion of the empirical data in subsequent chapters. The
theoretical background will be set by reviewing the Levelt Model
of speech production. It refines several former models (e.g.
Garrett, 1975; Shattuck‐Hufnagel, 1979) and provides, amongst
others, explanations for the preparation and production of
morphologically complex words. By being very explicit, it can
account for plenty of the phenomena reported in the literature – so
the hypotheses of the present dissertation are established within
this framework. Furthermore, the Interactive Activation (IA)
Model by Dell and the Independent Network (IN) Model by
Caramazza will be introduced and compared, focussing on the
representation of verb features and morphological processes. The
emphasis in discussing these models is placed on their coverage of
morphological complex words in the mental lexicon.
3.1 Lexical access and lexical selection
The language production process consists of composing a

preverbal message, coding it verbally, and articulating it. Thus,
Levelt (1989) suggests dividing speech production into three
levels: conceptualisation, formulation, and articulation. In the
literature, there is a broad consensus on the division into these
three general processing steps (cf. Levelt, 1989; Levelt, Roelofs &
Meyer, 1999; Dell, 1986; Caramazza, 1997). The process progresses
from concepts to perceivable code via word forms. The verbal
output can then be decoded again by a listener. Among other
47
48 CHAPTER 3: PSYCHOLINGUISTIC MODELS OF LANGUAGE PRODUCTION
issues, psycholinguistics is primarily concerned with the level of

formulation, i.e., of transforming conceptual into linguistic
structures.
The core process of language production is the lexical access.
This denotes the retrieval of lemmas and the corresponding word
forms from the mental lexicon that express the activated concept
(cf. Levelt, Schriefers, Vorberg, Meyer, Pechmann & Having,
1991). The division of lexical entries into lemma and word form is
an essential property of lexical access according to Levelt’s model
(but see Caramazza (1997) refuting such a bifurcation). The
division into these two stages has already been motivated in
Kempen & Huijbers (1983). Lemmas are linked to and activated by
concepts. A lemma (e.g. ‘sing‐’) is semantically specified and
makes syntactic information available (Levelt et al., 1999).
Roughly, it has a particular meaning and belongs to a syntactic
category, depending on which it is associated with further
syntactic information (e.g. gender for nouns, subcategorisation for
verbs); still, they are abstract entities because they do not contain
morphophonological information, which is stored on the word
form level. Crucially, a lemma can be realised in different word
forms which are instantiations systematically varying according to
their syntactic context. Lemmas “point” to word forms (Levelt et
al., 1999:187). Word forms are either auditory or written
realisations of lemmas. Some lemmas have alternating word
forms. This will play a crucial role in the course of this thesis.
While lemmas provide syntactic information, word forms carry
morphophonological properties (but see Caramazza, 1997).
Nothing is said about the representation of regularity of verbs.
In the 1990s, the idea of split lexical entries led to a debate
about whether lexical access proceeds in discrete or overlapping
stages, respectively. A lemma may either be selected prior to
morphophonological encoding (e.g. Levelt et al., 1999), i.e.
morphophonological information is activated only after the
selection of the correct lemma, or lexical selection and
morphophonological encoding proceed in parallel. This debate
affected the architecture of the models discussed below.
3.1 LEXICAL ACCESS AND LEXICAL SELECTION 49
Based on speech error data, Garrett (1975) proposed one of the

pioneering models – besides Fromkin (1973). He already
suggested two discrete stages of processing. Analyses of speech
error data revealed regularities and dissociations between word
and sound substitution errors. Corresponding to the distinction
between meaning (lemma) and form, word exchanges occur
between phrases and within the same word class, whereas sound
exchanges occur within a phrase and involve adjacent words in
the same phrase, i.e., sound exchanges often involve words
belonging to different word classes. Garrett attributed these
findings to two discrete serial stages which are not overlapping
but incremental (although multiple words for one phrase may be
processed in parallel on different levels). In competing models,
different types of information can be accessed simultaneously (e.g.
Dell, 1986). The selection of a lemma and its word form is
interactive. Hence, processing on both levels is overlapping.
Caramazza (1997) maintains that there is no separate lemma level,
and disputes successive retrieval of lemma and word form. He
argues for a simultaneous but also independent activation of
semantic, syntactic and phonological information (see 3.4 for a
discussion of the IN).
Selection of the lemma is preceded by a parallel check for the
fit between lemmas and the intended concept. There is consensus
about the initial activation of semantic alternatives on the lemma
level. With a combined picture naming and auditory lexical
decision task, Levelt et al. (1991) substantiate the view assuming
activation of multiple lemmas, i.e. co‐activation of lemmas
belonging to the same semantic field. Semantically neighbouring
words elicit robust priming effects as opposed to unrelated words.
Eventually, the target item receives most activation, and its lemma
is selected. This process is often labelled lexical selection.
However, there is a controversy regarding the number of
lemmas that become input for phonological encoding. Levelt
(1989) and Levelt and colleagues (1991) emphasise in an early
version of the model that lexical selection must converge on a
single lemma after parallel processing. Critical targets in the
experiment just mentioned (Levelt et al., 1991) were those items
phonologically related to semantic alternatives (like e.g. sheep –

(goat) ‐ goal). Compared to unrelated ones, they revealed no effect,
in accordance with the assumption of convergence on a single
lemma. This would support that only the selected lemma is
processed phonologically.
Yet, parallel activation of lemmas and word forms can easily be
witnessed in speech errors called blends. A blend is a merge of two
semantically related word forms like stummy for stomach and
tummy (from Fromkin, 1973). Roelofs (1992) noticed that most
blends involve near‐synonyms. If lemmas of two near‐synonyms
are equally active, both word forms would indeed be activated as
well.
While Levelt et al. (1991) could not demonstrate mediated
priming, Peterson & Savoy (1998) found phonological activation
of semantic alternatives (sheep primes goal via goat). These findings
were restricted to near‐synonyms. In more recent proposals (e.g.
Levelt et al., 1999), the authors take that special case of mediated
priming as support for a cascading model. Phonological properties
of a semantically processed but unselected item influence lexical
access. The assumption that only the selected lemma becomes
phonologically encoded remains held up. Near synonyms turn out
to be an exception, i.e. out of several activated lemmas, some
become phonologically prepared. The word form encoded first
will attain articulation. As the IA Model (Dell, 1986) is based on
compelling speech error data such as blends, it can naturally be
concluded that even slightly activated lemmas initiate
phonological encoding.
Two‐stage processing, as advocated by Levelt (Levelt, 1989;
Levelt et al., 1999), is implemented by the WEAVER++ model, a
comprehensive model of word form encoding (Roelofs, 1997). The
model can successfully produce words and even blends when
erroneously retrieving two lemmas.
All three language production models – the Levelt model, the
IA Model and the IN Model – conceive the mental lexicon as a
network that contains nodes on several levels. Nodes are
connected to each other both on their own level as well as across
levels. Selection of a target node depends on its activation, which
3.1 LEXICAL ACCESS AND LEXICAL SELECTION 51
varies gradually. The most activated lemma is selected. Selection

is more difficult in the presence of highly activated competing
nodes.
Activation processes are characterised by spreading activation,
another key concept in language production theories introduced by
Dell (1986). Every node with an activation higher than zero
immediately passes on (a portion of) activation to all connected nodes
(spreading). The activation of nodes increases with new activation
(summation) and decays over time if it does not receive new activation.
The activation of a selected node drops to zero immediately after
selection to avoid repeated selection of the same node.
The time required to retrieve lexical items is a function of the
frequency of the item and of morphological processes on higher
levels. Frequently used words are produced faster than less
frequently used words (Oldfield & Wingfield, 1965), as their
activation threshold is lower. In addition, Jescheniak & Levelt
(1994) specify that the frequency effect occurs on the word form
level. Jescheniak and Levelt had participants name pictures and
found that the frequency effects observed in picture naming
disappeared in a delayed cued‐naming paradigm. Because the
picture naming paradigm – with a sufficient delay – only
measures post‐lexical processes, they concluded that frequency
effects have their origin at the word form level. Otherwise, the
effects would not have vanished.
The ease of lexical access also depends on semantic factors. For
a long time, lexical access in the mental lexicon was seen as a
process of lexical competition because of the Stroop‐like
interference effects (cf. Schriefers, Meyer & Levelt, 1990; Levelt et
al., 1999): Naming of pictures is slower in the presence of
semantically related distractors (e.g. sheep ‐ goat) compared to
unrelated words (e.g. sheep ‐ tree). The idea of the lexical selection by
competition hypothesis is that activation percolates to semantically
related competitors, as the mental lexicon is conceived as a
network of nodes. Selection of the target is more difficult in the
presence of highly activated competing non‐target nodes. Any
semantically related distractor receives activation both from the
picture and from the distractor word. In contrast, unrelated words
are not affected by activation of the target, and therefore the target
itself is selected more easily.
Recently, a new perspective on semantic interference has
arisen: some studies showed semantic facilitation for semantically
related words.26 In carefully designed picture‐word interference
experiments, Mahon and colleagues (Mahon, Costa, Peterson,
Vargas & Caramazza, 2007) manipulated semantic distance
between target and distractor according to a standard graded
measure of semantic distance. The closer within‐category target‐
distractor pairs, the faster the target was produced. For example,
the picture of a horse has been named reliably faster with zebra
than with whale (Mahon et al., 2007). Consequently, the facilitation
effect cannot be lexical, and is not in accordance with the lexical
selection by competition hypothesis (for a review, see Mahon et al.,
2007). The data demand a reinterpretation of semantic interference
in lexical selection. The authors propose an explanation in terms
of increasing priming as the semantic distance decreases.
The second crucial finding by Mahon and colleagues on the
way to an alternative hypothesis was that articulation latencies
depended on criteria of the distractor relevant to the response, i.e.
semantic as well as syntactic information about the target was
helpful to exclude non‐target words from the response set.
Therefore, verbs did not interfere semantically with nouns, as they
were no potential targets.27 Participants were faster naming bed
26 The first discovery leading to a rethink was that low‐frequent distractors are not
facilitating, unlike high‐frequent ones (Miozzo & Caramazza, 2003). According to
the lexical selection by competition hypothesis, one would expect this because the
activation threshold of low‐frequent items is assumed to be higher, and they thus
should not interfere with the target as much as high‐frequent (almost pre‐
activated) distractors do.
27 In 2006, Finkbeiner and Caramazza attributed semantic facilitation effects to
response selection. In a picture‐word‐distractor experiment participants were
presented with semantically related/unrelated distractors that were either forward‐
or backward‐masked. Although unmasked distractors yielded semantic
interference, they showed reliable semantic facilitation under masked conditions,
i.e. facilitation due to semantic overlap of target and distractor, such as for the pair
car – truck as opposed to car – table. Masking prevents phonological activation, as
no competing responses need to be blocked. Together with semantic priming, it
results in facilitatory effects.
3.2 THE LEVELT MODEL (LEVELT, 1999) 53
with the distractor sleep than with shoot. Ultimately, Finkbeiner

and Caramazza (2006) and Mahon et al. (2007) propose the
response exclusion hypothesis, stating that unrelated distractors can
be suppressed sooner than items sharing the relevant properties.
In the field of lexical access, Finkbeiner and Caramazza (2006)
and Mahon et al. (2007) provide a new interesting empirical fact.
The new explanation licences the tentative conclusion that
semantic interference is not a lexical but rather a post–lexical
process.
3.2 The Levelt Model (Levelt, 1999)
3.2.1 Architecture
Levelt et al. (1999) designed a model of the production of single‐

word utterances. The production process of a single word is
assumed to consist of five levels (strata): (1) the lexical concept, (2)
the lemma, (3) the morpheme, (4) the phonological word and (5)
the phonetic‐gestural code. (2) and (3) are referred to as
grammatical encoding. Activation spreads from concepts to
articulation (unidirectionally), and there are no inhibitory
connections (see Figure 3). The following descriptive overview of
the model focuses on the stages and their output.
Figure 3. Levelt et al.’s (1999) Language Production Model.
In a first step, the content of the intended utterance is created. The

idea of the intended speech act is pre‐linguistically, conceptually
prepared by activating sensory specific mental representations. By
specifying semantics, lexical concepts (1) mediate between the
communicative goal and language‐specific words. Lexical
concepts are “concepts flagged by way of a verbal label.” (Levelt
et al., 1999:1). They constitute the output of the first stage. Because
of the organisation of mental representations in a network‐like

structure, lexical concepts of semantic associates become active
simultaneously. With CAT as target, ANIMAL and MOUSE will
be concurrently activated (see discussion above).
Levelt et al. (1999:4) argued that lexical concepts must be
represented holistically instead of being decomposed into
features. This organisation is necessary to avoid the hyperonym
problem (Levelt, 1989), i.e. automatic co‐activation of a
superordinate concept (e.g. animal) by the features of the concept
(e.g. dog) itself. If this was disregarded, the selection of a target
would be beyond the intention of the speaker, which is obviously
not the case.
From given lexical concepts, activation spreads to the lemma
level (2), where formulation of the message begins. The activation
spreads not only to the target lemma but also to the cohort of
semantically related lemmas. On the lemma level, grammatical
information is supposed to be stored in separate generic nodes
linked to the lemma, like for word category (Pechmann & Zerbst,
2002) and, with respect to verbs, subcategorisation information
(Branigan, Pickering, Liversedge, Stewart & Urbach, 1995) or
auxiliary type (Tabossi, Collina & Sanz, 2002). Generic means that
every entity with that particular attribute is linked to the
corresponding abstract feature representation.
The level of formulation is divided into two stages. The first is
grammatical encoding, comprising the retrieval of lemmas (2) and
the generation of a syntactic structure (Kempen & Hoenkamp,
1987). Lemmas are fully specified grammatically, but not yet
inflected. To retrieve grammatical properties, lemmas point to
their grammatical features. For morphologically complex words,
all properties are activated on the lemma level, but still as abstract
features (3). Having completed the retrieval of conceptual‐
semantic and syntactic information, the speaker has to bring all
morphemes into the correct order and find the appropriate
phonological word form (4) fitting the lemma (see the two‐stage
model discussion in section 3.1). Only selected lemmas activate
their phonological forms (Levelt et al., 1999:15). As with
grammatical features, lemmas also have pointers to word forms.
This second stage for retrieving the lemma’s word form is called
phonological encoding. Lemmas become phonologically specified,
implying the creation of a phonetic plan. For this purpose, the
ordered set of pointers brings phonemes into the correct serial
order, and phrases (utterances longer than one single word) are
structured and specified prosodically and metrically as well. Thus,
any successful outcome of the phonological encoding requires (a)
morphological, (b) phonological and (c) segmental processing (cf.
Levelt, 1999:5). Finally, the production process is completed with
the articulation (5) of the intended utterance. A self‐monitoring
mechanism applies from phonological words onwards and
ensures error‐free production. It observes internal speech during
the generation of words and utterances, and adjusts and corrects
the internal representations before articulation.
Chapters 6‐8 make use of the theoretical framework provided
by the revised Levelt Model. The model makes assumptions about
the representation and instantiation of tense, which serves as an
example for illustrating the processes involved in producing
inflectional categories here. Tense (event time) is in fact a
conceptual category (Levelt et al., 1999), but it has grammatical
reflexes in inflectional languages like German. It cannot be
realised by changing a single conceptual feature (e.g. ± past).
Rather, tense is expressed by complex words. Its instantiation in
inflectional languages has implications for the processes on the
formulation level, i.e. diacritic parameters need to be set to trigger
the appropriate morphological processes.
3.2.2 Diacritic parameters
Essential entities underlying production of complex words are

diacritics. Diacritics are connected to lemma nodes and can be
conceptualized as slots for specification of free parameters, such as
person, number, mood, or tense, which receive a value during the
process of grammatical encoding (Levelt et al., 1999:6). Setting
diacritics at the lemma level is necessary for the encoding of
appropriately inflected word forms.
As Levelt et al. (1999) is primarily a theory about very simple

word naming, there is only some hand waving about tense and its
information flow from conceptualization to setting the diacritics.
However, Janssen (1999:46) assumes a distinction between concept
nodes and concept classification nodes at the conceptual level,
where the message is created. Lexical concept nodes point and
refer to objects or actions. Conceptual processing of inflectional
categories like tense or number is separated from the lexical
content and represented by concept classification nodes. Concept
classification nodes carry variable information that cannot be
stored fixed to lexical concepts. For instance, TABLE can be
modified by PLURAL to TABLES. Hence, concept classification
nodes modify and extend the meaning of lexical concept nodes, and
are not interpretable in isolation.
The division into concept nodes and concept classification nodes
ensures independence of lexical processing from inflectional
processing. Inflectional morphemes are indeed separable from
lexical content, as is known from speech errors like morpheme
stranding, extensively studied by Garrett (1982).
The information on the conceptual level is transferred to the
formulation level through spreading activation. The lemmas then
contain the current value of the inflectional categories in the form
of diacritics, i.e. the diacritics temporarily store a word’s variable
information like e.g. present or plural. The values depend on the
context and concept of the utterance and have their origin at the
conceptual level. Further processing of inflectional morphemes is
set and determined by diacritics. Later in processing, diacritics are
represented by inflectional affixes.
The diacritic feature of tense involves conceptualization of
temporal reference, such as +past or +present. It is noteworthy that
tense diacritic parameters are essentially identical for
morphologically regular and non‐regular verbs. Other diacritic
parameters, such as verb number for subject–verb agreement
(Bock, 2004) need not be specified during formulation. They are
assumed to operate only during syntactic encoding. Diacritics
differ from syntactic properties in the latter’s permanent link to
the lemma via generic nodes. Diacritics carry information that is
copied prior to the lemma retrieval from the concept classification

nodes to the diacritics of the currently activated lemma.
Faroqi‐Shah & Thompson (2007) provide evidence for diacritic
encoding from neuropsychological data. Agrammatic patients
exhibited verb inflection difficulties. Their encoding of verb
inflection was examined with a multiple‐choice sentence
completion task. Faroqi‐Shah & Thompson (2007) report that
errors in the production of verbs were likely to occur when
temporal reference was involved. Agrammatic aphasic patients
may have had difficulties with diacritic encoding and retrieval of
tense features rather than with inflection per se, i.e., their errors
were not the direct result of morphological complexity but of
impaired setting of diacritics. A temporal context was provided
for all trials. However, it cannot be excluded that the deficit in
tensed verbs is due to a failure in conceptualisation of tense.
The Levelt Model’s three most important characteristics are (1)
seriality, (2) modularity and (3) feed‐forward activation. The
levels function autonomously in a hierarchical order and pass
down their output to the lower levels (see Figure 3), but not until
all processing is completed. There is no feedback between the
levels as in former models (e.g. Fromkin, 1973; Garrett, 1975;
Shattuck‐Hufnagel, 1979). Modularity calls for an additional
assumption: incrementality. Longer utterances are not prepared as
a whole but are split up into shorter units so that some parts can
be processed on a higher level while subsequent parts are already
simultaneously processed on lower levels. Yet, internally, each
processing step is still serial. This ensures fluency, especially for
longer utterances.
3.3 The Interactive Activation Model
The Interactive Activation Model is a theory of sentence

production, as explicitly stated in the introduction to Dell (1986).
The fundamental difference between the Levelt Model and the
Interactive Activation (IA) Model is the activation flow. While the
Levelt Model is organised into strictly serial modules, the IA
3.3 THE INTERACTIVE ACTIVATION MODEL 59
Model is highly interactive, i.e., activation spreads not only

unidirectionally top-down but also bottom-up. This is a prerequisite
for feedback processes, although the feedback is restricted to
adjacent levels (local). Another basic assumption of the IA Model
by Dell (1986) is that connections are excitatory rather than
inhibitory.
Most of Dell’s processing hypotheses stem from speech error
analyses. He observed that speech errors result in real words of
the language more frequently than random chance would predict
(lexical bias effect, Dell & Reich, 1981). Dell interprets this finding as
a consequence of feedback from activated word forms to the
lemmas via backward spreading. This entails that lemma selection
and phonological encoding are interrelated through bidirectional
connections. Crucially, semantic alternatives of a target become
phonologically activated as well. If errors in phonological
encoding activate existing word forms (like rude lip / lewd rip),
erroneously activated word forms induce activation in their
corresponding lemmas, which in turn add to the activation of the
word form again. Consequently, such non‐target word forms
reach the activation threshold quickly enough to be selected. It has
also been shown in experiments28 that errors occur less often if the
resulting form is not a real word (like luke risk / ruke lisk).
Dell’s (Dell & Reich, 1981) second observation is the relatively
frequent occurrence of mixed errors, that is, the production of
linguistic items which are both semantically and phonologically
misguided, e.g. producing rat instead of cat, which is semantically
and phonologically related. For reasons of logical consistency,
these can only arise from simultaneous activation of semantic and
phonological information in parallel. With this in mind, it appears
implausible to opt for a model consisting of two discrete stages.
Hence, Dell & Reich (1981) take this finding as evidence for
28 Speech error induction techniques include, for example, exposing participants to

an auditory distractor word immediately before the utterance of a target word, or
increasing the speech rate. Elicited slips of the tongue can be caused by phoneme
exchanges resulting in anticipations, and perseverations.
upward feedback processes from phonological encoding to the

lemma level.
Last but not least, and in a similar line of thought, interactivity
is the explanation for the word superiority effect (Dell & Reich,
1981). Readers recognise letters in the context of a word better
than when the same letters are presented in pseudowords or in
isolation (Reicher, 1969). The word superiority effect shows that
speech processing does not only involve bottom‐up but also top‐
down processes.
Dell emphasises that errors are not faults of the cognitive
system, but must be expected and accommodated by the theory’s
assumptions. The architecture of the IA Model can successfully
simulate production processes, but it can equally model
performance leading to speech errors. Inappropriate items are
selected simply because of their higher activation. Therefore,
errors are a product of random noise, the emergence of which is
integral to the model and presumably also to the brain.
Having outlined the general principles of the Interactive
Activation Model, it is necessary to say a few words about its
architecture, as it will prove relevant to the subsequent more
detailed discussion of individual processing steps, especially
morphological encoding. The lexicon is organised as a structured
network. It consists of nodes on the conceptual, lexical (i.e. lemma)
and phonological levels. Concepts are represented feature by
feature in a decomposed form with all concept nodes
interconnected.
If a node on the conceptual level is activated, it activates all
connected lexical nodes on the lemma level. Every activated lexical
node spreads activation to the phonological level (in contrast to
Levelt, 1989; Levelt et al., 1999, where phonological activation is
strictly limited to the selected lemma). The lemma with the
highest activation will finally be phonologically represented and
articulated. But in the IA Model, activation spreads back to upper
levels, reflecting its interactive character.
The levels mentioned above contain declarative knowledge
about:
3.4 THE INDEPENDENT NETWORK MODEL 61
(1) concepts
(2) words/morphemes
(3) syllables
(4) phonemes
(5) phonetic features
Word nodes are marked for syntactic and morphological category.

Representations are constructed by generative rules stored on
each level, i.e. by combinatorial knowledge for syntactic,
morphological and phonological encoding. The interplay between
lexicon and rules follows the slots‐and‐fillers‐principle: rules
specify the item for the appropriate slot and thus compose the
constituent lemmas, morphemes and phonemes into a coherent
sequence.
The combination of rules and lexicon creates an opportunity
for modelling the production of regular and non‐regular verbs, in
that morphological frames may include slots for verb stems and
affixes.
To sum up, the main difference between the Interactive
Activation Model and the Levelt Model is the feedback, and
thereby the spreading of information and activation from lower‐
level to higher‐level processes. As Dell’s model is not a serial
model, it is possible that the phonological form influences the
selection of the lemma. Phonological encoding begins before the
final selection of the lemma. Hence, it allows for a short period of
time in which both lemma and word form are activated.
Furthermore, all activated lemmas (semantic competitors) become
phonologically encoded.
3.4 The Independent Network Model
The models introduced in the preceding sections have in common

that they represent lexical information in the form of networks,
and that processes are described as spreading activation between
nodes. These characteristics are no less essential for the third
model – the Independent Network Model. Nevertheless, it
presents a challenging alternative to the two‐stage theory with

respect to activation of syntactic features. Strikingly, Caramazza &
Miozzo (1997) and Miozzo & Caramazza (1997) have found that
Italian speakers can retrieve partial phonological information
about the target without knowing its grammatical gender.
Levelt constructed his model mainly based on reaction time
data, while Dell’s model is based on speech error analyses. For his
Independent Network Model, Caramazza (1997) used empirical
data gathered from cognitive neuropsychology and linguistic
performance of aphasic patients. He discovered that some anomic
patients can provide information about grammatical properties of
words even if they cannot produce the word itself. Conversely,
there are also patients who are able to produce the phonological
form but do not have access to grammatical information about the
words they are producing – an observation akin to the
dissociations between semantic and syntactic information found
by Caramazza & Miozzo (1997) and Miozzo & Caramazza (1997).
For these reasons, Caramazza (1997) addresses the relationship
between semantics, syntax and word form in a novel way. He
proposes that information is stored in three independent networks
– representing semantic, phonological, and syntactic knowledge,
respectively.
1. The lexical‐semantic network stores the meaning of words,

represented as features.
2. The lexical‐syntactic network contains grammatical
information (e.g. gender).
3. The phonological network stores phonological output
representations (lexemes).
The networks of the IN Model are autonomous. Information from

each network can be retrieved independently; phonological
lexemes, for instance, can be accessed without prior retrieval of
syntactic information.
The IN Model also has a network containing orthographic
output lexemes for written language production. Caramazza
considered the modality of production to be essential because the
3.4 THE INDEPENDENT NETWORK MODEL 63
architecture of the IN Model is based on data from patients with

modality‐specific impairments: having observed patients like SJD
and HW (Caramazza & Hillis, 1991), who exhibited selective
grammatical deficits in written and oral production, respectively,
Caramazza and colleagues concluded that phonological and
orthographic word forms must be stored separately. The output
network has to be modality‐specific. All word forms are, however,
connected to the same semantic network, where concepts are
stored in a decomposed way. Both orthographic and phonological
word forms are connected to the syntactic network, where
grammatical properties such as word class,29 gender, and tense are
represented. Grammatical features are organised into
subnetworks; a German subnetwork for gender would, for
example, consist of nodes for masculine, feminine and neuter.
Links between nodes of a subnetwork are inhibitory because no
more than one can be demanded at a time.
In the IN Model, the modality‐specific word forms can either
be accessed via activation of grammatical features or without prior
activation of the syntactic network (e.g. tip‐of‐the‐tongue (TOT‐)
states30), since the networks are independent. Unlike in the
previous models (the Levelt Model and the Interactive Activation
Model), activation is passed on instantaneously from the lexical‐
semantic network to the word form networks, with feed‐forward
activation only. The IN Model does not feature a modality‐
independent lemma level.31 Links between lexical‐semantic
representations and phonological/orthographic word forms are
direct and require no intermediary. Consequently, all word forms
sharing semantic features become activated simultaneously,
similar to the Interactive Activation Model. Caramazza arrived at
29 Word class might be represented in the lexical semantic network because it has
a “semantic reflex”, i.e. semantic attributes can determine word class like for
example “objecthood” (Caramazza & Miozzo, 1997: 340).
30 Tip‐of‐the‐tongue phenomena are states in which the phonological form of
target words is temporarily not or only partially available, while syntactic
information is preserved and can be accessed. This has often been taken as
evidence for the lemma‐word form distinction.
31 Caramazza considers lemma selection and activation of grammatical features as
one single process and therefore proposes to fuse the two in the model.
this assumption after he observed patients producing different

semantic errors in naming and writing for the same target. Orally,
patient PW produced pliers instead of tweezers and wrote needle in
response to the same picture of tweezers. He then named it pliers
again (cf. Caramazza, 1997). The explanation for these contrasting
error patterns in both output modalities successfully avoids
reference to abstract lemma nodes and instead rests solely upon
activation spreading between lexical representations and word
forms (forward spreading only). Nonetheless, this argument,
which relies solely on patient performance, is not incontrovertible,
since the data might as well be explained in a model assuming
lemma nodes.
Despite this objection, dispensability of the lemma level is not
in conflict with Levelt (1989) and Levelt et al. (1999). Caramazza’s
data rather confirm the two‐stage theory, as Levelt et al. have
never claimed that the grammatical features must necessarily be
selected and accessed with each lexical access. Activation without
selection typically has different consequences than activation with
selection. If grammatical features need not be retrieved in a
syntactic context, selection may easily be skipped.
In Caramazza’s model, tense is represented as a subnetwork of
the grammatical network. The theory is not very explicit about
this, except for stating that tense nodes receive activation from the
lexical‐semantic network. Caramazza explicates that tense is an
extrinsic feature opposed to intrinsic features like gender (cf.
Bordag & Pechmann, 2009). Extrinsic features are derived from
the conceptual level. The value for tense (e.g. present or past) is
determined by its use depending on the context (Caramazza,
1997:186; footnote 4). Unfortunately, the IN Model does not make
any claims about morphological processing at all. Hence, the
processing of tensed verbs which are morphologically complex
cannot be modelled. Existing data and discussion concerning this
issue are neglected by the author.
3.5 DISCRETE VERSUS CASCADED PROCESSING 65
3.5 Discrete versus cascaded processing
An issue that is a matter of debate among theories of word

production is how activation flows between the two levels of
lexical representation. Discrete serial models of lexical access (e.g.
Levelt, 1989; Levelt et al., 1999) advocate activation flow strictly
from layer to layer. Although multiple lexical nodes become
activated, only one node (the selected one) will spread activation
from the semantic to the phonological layer. Phonological
encoding starts only after the lexical selection is completed. In
contrast, some activation spreading theories of lexical access, such
as Humphreys, Riddock, and Quinlan's (1988), propose that
activation flows continuously from layer to layer in a cascaded
fashion. Activated nodes spread some proportional activation
regardless of which node will eventually be selected. Phonological
information will not only be activated for the selected node.
Earlier experimental studies spoke for discrete processing
(Schriefers et al., 1990; Levelt et al., 1991). More recent studies
supported cascading models (Peterson & Savoy, 1998).
Humphreys et al. (1988) is a cascade model assuming only
forward spreading of activation. Activation does not spread from
the phonological level back to lexical levels. This is an important
difference to cascading interactive models which also allow
backward spreading (e.g. Dell, 1986). Models with backward
spreading can account for mixed errors and the lexical bias effect.
3.6 Remarks on diversity of models
All three models – the Levelt Model, the Interactive Activation

Model and the Independent Network Model ‐ essentially capture
the lexical retrieval process in the form of spreading activation,
but they make different claims about the relationships between
the activation of syntactic, semantic and phonological information.
The models reviewed above are still under evaluation and, of
course, find themselves in a process of constant re‐consideration.
As of yet, there has neither been conclusive empirical evidence nor
a conclusive argument in exclusive favour of either of the models.

Cascade models, however, are more difficult to test because when
faced with the potentially problematic finding of a lack of
phonological activation of unselected items, they could
hypothesize that the activation exists but is too weak to be
detected.
While the Interactive Activation Model above constitutes a
more comprehensive model of speaking, including both word and
sentential utterances, the present study will focus on aspects of
single word production. Specifically, the Levelt model's
postulations about word‐form encoding processes are scrutinised
in order to determine whether they remain viable when producing
regular and non‐regular verbs. Moreover, Levelt et al. (1999)
present the most explicit model of production of morphologically
complex words.
3.7 Producing morphologically complex words
The beginning of this chapter introduced the Levelt model as one

example of speech production models. This chapter will address
the Levelt Model from a procedural perspective. How can the
generation of complex words, especially of tensed verbs, be
explained in the architecture of the Levelt model? I will adhere to
the description of generating the verb form escorting ‐ a
morphologically complex verb – as provided in Levelt et al. (1999),
and primarily focus on grammatical encoding.
Let us take the lemma escort with the corresponding set of
pointers on the lemma level (Levelt et al., 1999) as a starting point.
First, the lemma is activated, and its diacritic parameters are set
for progressive. Subsequently, Levelt postulates that the
corresponding morphemes are retrieved from the form lexicon.
The lemma points and spreads activation to the morphemes escort
and -ing on the form level. They are then brought into the correct
serial order and concatenated.
As has already been noted in 3.2, word form generation requires
the retrieval of the lexical stem plus three types of information:
3.7 PRODUCING MORPHOLOGICALLY COMPLEX WORDS 67
morphological, phonological, and segmental details of the intended

word form. Every required morpheme is immediately retrieved
from the mental lexicon. Word forms for activated morphemes
contain metrical and segmental information. Metrical information
about morphemes includes whether morphemes are free or bound:
-ing for example, cannot stand alone. Morphemes contain a
specification of whether they are phonological heads (like escort) or
not (like -ing). In addition, they are numbered for ordering, so that
concatenation can proceed successfully. Afterwards, activated
phonological plans can be modified by syllabificational processes to
ensure correct pronunciation: es.cor.ting (instead of e.scort.ing, for
instance). Syllabification varies depending on the phonological
environment and results in different phonological words.
Phonological words, in Levelt’s understanding (Levelt et al., 1999),
are a prosodic unit defined in terms of syllabification prior to
articulation. They differ from the phonological plans for targets
because they are adapted to the context like in es.cor.ting where /t/
becomes onset of the third syllable. Differently with popart which
consists of two phonological words: pop.art. Finally, phonological
words will be encoded phonetically, and incrementally converted
into muscle contractions and sound waves.
For complex as well as for mono‐morphemic words, diacritics
have to be set at the early stage of conceptualisation, i.e.
independently of the complexity of the target. The encoding
process then proceeds with selecting or generating that word form
which corresponds to these diacritics. Diacritics for person and
number agreement between subject and verb are extrinsic features
that can be set prior to grammatical encoding. This information is
available upon encountering the subject, and does not depend on
verb lemma retrieval.
Furthermore, setting diacritic parameters for tense is
independent of the regularity status of the verb. The difference
between regular and non‐regular inflection arises only in later
processing stages: In the case of non‐regular inflection, there is no
one‐to‐one mapping between diacritics and affixes. Non‐regular
verbs select an allomorph of the stem. Selection of an inflectional
affix is determined by the inherent grammatical properties of a

lemma and by variable diacritic parameters.
To retrieve regular words, Levelt et al. (1999:12) point out that
two nodes are involved only at the word form level, one for escort
and the other one for -ing. “Regular inflections are probably all of
this type, but non‐regular verb inflections are not, usually. The
lemma go-past will activate the one [emph. HT] morpheme went.”
For both types of inflection, the first step is selecting and setting
diacritics. Production then proceeds with selecting or generating
the word form that corresponds to the diacritics.
For concatenation of morphemes to fully inflected word forms,
Levelt et al. (1999) assume rules to apply. That is, rules definitely
apply for processing regular words, and apply occasionally if
required for non‐regulars. These rules are implicit in the Levelt
model. They are more explicit in the Words and Rules Theory by
Pinker (1999) and Clahsen (1999).
3.8 Morphological processing in comprehension
Comprehension is the other side of the coin when investigating

the processing of morphologically complex words. Although there
are fundamental differences between comprehending and
producing language, research from both modalities addresses the
organisation of the mental lexicon (for a study on the regular/non‐
regular distinction in comprehension see Baayen, Dijkstra &
Schreuder (1997)). Both modalities use different representations of
word forms (Zwitserlood, 1994), but it is likely that they share
conceptual representations – everything else would be
computationally inefficient (Cutting, 1998).
One of the first and most influencing models of word
recognition has been provided by Taft and Forster (1975).
Morphological analysis and segmentation are integral parts of
lexical storage and retrieval of prefixed words respectively,
together with separate representations of base morphemes (i.e.
stems) and affixes. According to Taft and Forster (1975), lexical
access of morphologically complex words means stripping affixes
3.8 MORPHOLOGICAL PROCESSING IN COMPREHENSION 69
and accessing constituent morphemes individually even in the

case of bound morphemes (for example, ‐juvenate in rejuvenate).
Thus when a prefixed word is to be recognised, it is first
decomposed into its prefix and stem, and lexical access then
proceeds on the basis of the stem only. The authors present
evidence from a lexical decision task, where reaction times to
items with bound morphemes were longer than for pseudo‐bound
control words (like ‐pertoire in repertoire). Therefore, it was
suggested that affix stripping is a highly automatic process, and
the delayed reactions with bound morphemes indicate the need
for reanalysis if no matching stem can be found in the lexicon.
The model Taft and Forster put forward concerns both the
processing of visually recognised prefixed words and their storage
in the mental lexicon. A prefixed word is stored in the lexicon as a
representation of its stem. The lexical entry includes information
about which prefixes can combine with the stem to form a word.
For example, Taft and Forster interpret lexical entries as having an
internal structure like RE(JUVENATE), hence, complexity is
represented in lexical entries. Prefix information is stored within
this lexical entry. The authors exemplify this point themselves:
“the theory that states that the entry for rejuvenate is re(juvenate)
would claim that admit, remit, and so on, all have separate lexical
entries.” (Taft & Forster, 1975:645)
The ideas of Taft and Forster (1975) were revised by Caramazza

and colleagues, who developed another model of word
recognition: The Augmented Addressed Morphology Model
(AAM) (Caramazza, Miceli, Silveri & Laudanna, 1985; Caramazza
et al., 1988) presents an elaborate model of the lexicon which
clearly incorporates the morphological structure of lexical items
and morphological procedures for lexical access. Taft and Forster
(1975) did not assume holistic processing of words at all. The
AAM, however, unifies decomposition and full‐listing.
Word form representations and processing, according to AAM,
are modality specific. Basically, processing of written words relies on
what the stimulus carries in its surface: orthographic information.
Moreover, word recognition follows the principle of similarity: The

incoming stimulus is checked against stored word forms. A letter
string activates its whole word representation (if there is any) as well
as its morphemes. For example walked activates walked + walk- + -ed.
Also word forms of similar words will be activated, e.g. walks or
walking. The orthographic representation which gains the most
activation and therefore reaches a threshold first will activate its
lexical entry. The two processing mechanisms – full‐listing and
decomposition ‐ work in parallel. However, it is assumed that full‐
listing delivers the output more rapidly than the activation of several
constituting morphemes. A stimulus is represented as a full form if it
has been learnt by prior exposure. Otherwise the word is accessed by
the lexical representations of its morphemes.
In Caramazza et al.’s (1988) empirical work, reactions to non‐
decomposable non‐words (canzovi) were fastest, followed by
reactions to non‐words with partial morphological structure
(canzevi). Most difficulties and longest reactions times arose with
morphologically legitimate non‐words (cantevi) because they were
parsed as if they were morphologically complex. The parser was
slowed down by the identified constituent morphemes. Therefore,
non‐words with suffixes were read using a morphological parsing
address procedure. Materials of the lexical decision experiment in
their study of Italian were controlled for frequency and similarity,
and they tested different types of non‐words.
Summing up, the lexicon contains full‐form representations of
word forms and decomposed word forms of roots and affixes.
Whole word processing is assumed for known words and
decomposed processing for unknown and regular words. As a
consequence of orthographic forms guiding processing, all
irregular words must be stored because they do not have a trigger
for decomposition on their surface. Roots and affixes are
independently represented but regular (“major”) roots linked to
corresponding suffixes. For the Italian verb correre [to run], corr‐ is
a major stem linked to all suffixes (of which some are linked
through inhibitory connections), and cors‐ is linked only to
suffixes it can co‐occur with.
4 Representation and processing
of grammatical features
In investigating verb and noun inflection, linguistic units are
divided into smaller units, i.e. the noun or verb stem and
grammatical features such as gender, number, person or tense.
These grammatical features provide valuable clues for structural
processes in language comprehension as well as production. Most
production models agree upon abstract representation of
grammatical features on the lemma level (e.g. Levelt, 1989; Levelt
et al., 1999; Dell, 1986). This information is assumed to be generic,
i.e. all lexical items sharing a particular property are linked to the
same abstract feature. In the following, the structure of the mental
lexicon as well as of lexical entries is outlined before theoretical
approaches on the classification of grammatical features are
discussed. Finally, the processing of two types of grammatical
features, namely gender and declension/conjugation class, is
reviewed in more detail.
4.1 Representation of lexical information in the

mental lexicon
Theories of language production (and comprehension) typically

propose that word retrieval involves the selection of lexical
information. The mental lexicon is the repository which is
assumed to represent lexical information and is part of the human
long‐term memory. In which form linguistic information is stored
is a question of the configuration of the mental lexicon. The
proposals vary to a great extent for particular requirements of
semantic, syntactic, morphological and phonological information
71
72 CHAPTER 4: REPRESENTATION AND PROCESSING OF GRAMMATICAL FEATURES
(for a review see Rapp & Golrick, 2006). Conventionally, the

mental lexicon organises speakers’ vocabulary in the mind like in
a dictionary. It is conceived of as a list of words, the lexical entries
(cf. Pustejovsky, 1996). There is controversy about the content of
lexical entries. Agreement has been reached regarding information
about (a) the pronunciation, (b) the meaning, (c) morphological
properties, and (d) syntactic properties of its entries32 (Levelt,
1989:182). Keeping as close as possible to the definition given by
Levelt et al. (1999),
“…a lexical entry is an item in the mental lexicon, consisting of a

lemma, its lexical concept (if any), and its morphemes (one or more)
with their segmental and metrical properties.” (Levelt, 1989:182)
Several recent directions in research are pointing to a view of the

mental lexicon that is not that of the classical dictionary (e.g.
Elman, 2004). Rather, the lexicon is viewed as a complex, dynamic
system of knowledge in which information is not stored passively
but actively directs on mental states.
4.1.1 Structure of the mental lexicon
The fundamental question about the organisation of the mental

lexicon is: What is the primary unit of representation? The answer
concerning the exact type of linguistic representations is an issue
of the linguistic level addressed. Theories about for example
semantic representations and morphological processing provide
conflicting answers to this question.
How semantic knowledge is organized is still an unresolved
issue (cf. Aitchison, 1997). In the literature are two main accounts
that specify how semantic knowledge is represented and
retrieved. The Feature Comparison Model (Smith, Shoben & Rips,
1974) assumes semantic knowledge to be decomposed in features.
Concepts share features and their meaning is computed during
32 Idiosyncratic information alike

4.1 REPRESENTATION OF LEXICAL INFORMATION IN THE MENTAL LEXICON 73
processing. The more features two concepts have in common the

more similar are the concepts. On the other hand, the Hierarchical
Network Model (Collins & Loftus, 1975) describes semantic
knowledge as categorical, associative relations among concepts in
a network‐like structure. The closer and the stronger the
connections are between two items the more similar are the two
concepts. This structure reflects hierarchies and hyperonyms
inherit their properties to hyponyms.
Of more interest for the current study are theories about
representation and processing of morphologically complex and
simple words. Aspects of morphology in the mental lexicon could
enable us to examine the roles and interplay of storage and
computation in the mind. Representation of language is an
important component of a theory of language processing. It is not
easy to distinguish representation and rules and to keep them
apart.
Mainly two ideas about the representation of morphologically
complex words are pitted against each other in the literature. First,
Butterworth (1983) claims non‐compositional full listing of all
words in the lexicon. There are lexical entries for each word forms
like for kiss, kissed, kissing etc. However, all the entries include a
representation of their morphological structure. A variant of the
full listing assumption was proposed by Lukatela and colleagues
(Lukatela, Gligorijevic, Kostic & Turvey, 1980) arguing for a
satellites entries model. The lexical organization of the inflected
nouns is assumed with the nominative singular as base form in
the centre (nucleus) connected to all possible inflected forms as
satellites grouped around it. In a lexical decision task, participants’
reaction times were fastest in response to the nominative singular
and did not differ for genitive and instrumental. Lukatela et al.
(1980) argue against morphological decomposition.
Secondly, the competing position reckons representation of
separated constituting morphemes in the lexicon, i.e. only stems
like kiss- plus representations for exceptions (e.g. sheep) are listed
(Taft, 1981; Taft & Forster, 1975). In these models, the basic
operation is affix‐stripping and the remaining root is looked up in
the lexicon. Facilitation effects by morphologically related
distractors in picture naming tasks are a good argument for

decomposed storage in the mental lexicon (Dohmes, Zwitserlood
& Bölte, 2004, among others). Intermediate positions between full
listing and decomposition by Pinker (1999) for production and
Caramazza et al. (1988) for comprehension have already been
described in sections 2.3 and 3.7. These conflicting proposals still
form an inconclusive picture, especially for the representation of
non‐regular forms.
4.1.2 Underspecified lexical entries
After the overview of accounts about the structure of the mental

lexicon regarding semantic and morphological information this
section continues with an approach about the structure of lexical
entries as parts of the mental lexicon by Clahsen (1999) who
argues for underspecified lexical entries.
Arguing persuasively for Dual Route processing, Clahsen
(1999) and Clahsen et al. (2001, 2002) propose structured,
underspecified lexical entries for non‐regular participles such as
(ge)trunken [drunk]. The lexical entry of non‐regular verbs has two
layers: the verb stem on top of the lexical entry is an unmarked
mother node. Information about the ending and about ablaut is
coded by subnodes spreading from each corresponding unmarked
mother node. Subnodes are feature pairs formed upon the pattern
<phonological string, morphological feature value>. Non‐regular
forms conceived of as subnodes of lexical entries get features from
the mother node by inheritance (Clahsen, 1999). The concept of
underspecified lexical entries is based on Minimalist Morphology
(Wunderlich, 1996). The past tense and participle forms of non‐
regular verbs are pretty similar to their stems. Often they differ
only in length and vowel quality. Underspecified lexical entries
are an economic way of representation and can account for the
high degree of similarities of non‐regular verbs (see Figure 4).
Lexical entries of non‐regular verbs are schemes that restrict
the possible non‐regularities. In contrast, entries for regular verbs
consist only of the mother node. Participles are built by regular
affixation of the affix -t to the mother.
4.2 INTERNAL AND EXTERNAL FEATURES 75
Figure 4. Clahsen’s (1999) structured lexical entry for the non‐regular

verb to drink.
4.2 Internal and external features
Turning back to language production, retrieving word forms

entails also the retrieval of grammatical features. On the level of
grammatical features the discussion is very close to the primary
unit of representation in the mental lexicon. In most production
models, grammatical information is stored on the lemma level in
form of generic nodes.
The grammatical features can be classified as internal or external
(Bordag & Pechmann, 2009). The lexical information of internal
features only becomes available with the activation of a specific
lemma (e.g. grammatical gender). External features, however, are
conceptually determined by the context and therefore are
available before any lemma is activated. Their value is variable
and lemma‐independent (e.g. number or tense). A similar
distinction between extrinsic (contextually determined) and
intrinsic (inherently associated) features was made first by
Caramazza (1997) and Schiller & Caramazza (2002). The definition
of internal and external by Bordag and Pechmann (2009) displays a
strong dependency on the notion of the lemma and its activation –
a concept not found in Caramazza’s model. According to Bordag
and Pechmann, external features are obligatorily activated and
specified as diacritic parameters during grammatical encoding in
the Levelt Model. If external values specify the output sufficiently
and internal feature values are not distinctive for inflectional

processes, internal features can be bypassed33.
With regard to the grammatical features, gender and
declension/conjugation class, their values remain unaffected by
agreement or government relations. They are both internal
features and stated arbitrarily in the form of generic features
which has been revealed by several picture‐word interference
(PWI) studies on grammatical feature activation (Schriefers, 1993;
Bordag & Pechmann, 2009; see below).
In PWI studies, a presented picture (target) has to be named
whereby the articulation latency is measured. The picture
(“target”) emerges either simultaneously, or somewhat earlier or
later (different stimulus onset asynchronies (SOAs)) than an auditory
or written distractor34. Although participants are asked to ignore
them, distractors are processed unconsciously and the type of
distractor has an impact on the picture naming. Usually,
semantically related distractors cause interference while
phonologically related ones lead to facilitation (Schriefers et al.,
1990). Moreover, variation of SOAs is an important variable for
obtaining picture‐word interference effects and studying the time
course of encoding. Schriefers et al. (1990) presented auditory
distractors at varying SOAs. If lexicalization proceeds in two
distinct stages such that lexical selection precedes phonological
encoding, one would predict semantic interference and
phonological facilitation to occur at distinct SOAs. The results
supported the idea of two distinct stages, because the time to
name target pictures was affected by the semantic distractors at
short SOAs only. In contrast, an influence of the phonological
distractors was only present at long SOAs.
33 An exception is the gender congruency effect that is observable for bare noun
naming. Other accounts considering the bypassing of feature selection are
Schriefers (1993) and Levelt et al. (1999).
34 Following conventions in the literature, “targets” and “utterances of
participants” are quoted, distractor words are underlined, everything else
(morphemes, lemmas…) is written in italics.
4.3 PROCESSING GRAMMATICAL GENDER 77
4.3 Processing grammatical gender
The PWI paradigm was used first by Schriefers (1993) for

examining the selection of grammatical gender in language
production. In this study Dutch participants were asked to name
depicted coloured objects that were paired either with gender
congruent or gender incongruent distractor words. Note that in
Dutch common gender (de groene stoel [the green chair]) is
distinguished from neutral gender (het groene bed [the green bed]).
The stimuli should be named with noun phrases consisting either
of determiner + adjective + noun (e.g. het groene bed [the green
bed]) or adjective + noun (e.g. groen bed [green bed]). Dutch is a
gender‐marked language in which adjectives are gender‐inflected
only when determiners are absent. The study showed that
pictures were named faster when the name of the target picture
and a distractor noun were gender congruent. According to
Schriefers (1993), delayed reaction times in incongruent conditions
reflect a competition of two different gender feature values for the
final selection. The distractor word automatically activates its
corresponding gender value and distracts the gender selection for
the target. Usually, gender is proposed to be represented as
abstract generic nodes independently of semantics (Jescheniak &
Levelt, 1994; among many others; see Figure 5).
Figure 5. Lexical information on the lemma level is represented as generic

nodes. paard and hond are linked to the single noun‐node (taken from
Jescheniak & Levelt, 1994).
The gender congruency effect has been replicated by Van Berkum

(1997), La Heij and colleagues (La Heij, Mak, Sander &
Willeboordse, 1998) and, for German, by Schriefers and Teruel
(2000). However, it could not be verified for French and Italian
(Miozzo & Caramazza, 1999; Alario & Caramazza, 2002). Hence,
this effect seems to be language‐specific. In Romance languages,
the selection of the determiner often does not only depend on the
gender of the noun phrase (i.e. grammatical features) but also on
the phonological context (i.e. the onset of the word following the
determiner). The Italian definite determiner il (masculine singular,
e.g. il tavolo [the table]) is replaced by lo if the following word
4.4 PROCESSING DECLENSION AND CONJUGATION CLASSES 79
begins with a vowel, an affricate, or consonant clusters with /s/ or

/gn/ (e.g. lo strano tavolo [the strange table]). While classically, the
gender congruency effect was attributed to the lemma level,
Caramazza and his colleagues (Miozzo & Caramazza, 1999; Alario
& Caramazza, 2002; Schiller & Caramazza, 2003) instead perceive
it as a competition of determiners instead of abstract gender
nodes. The determiner cannot be selected until sufficient
phonological context is available. Effects resting on grammatical
feature selection would not be detectable during later processing
stages. Caramazza and colleagues conclude that the observed
congruency effect is a determiner effect rather than a lemma effect.
4.4 Processing declension and conjugation

classes
Research on language production has mainly focused on noun

and gender processing, so that only little is known about the
processing of verbs. This is because most of the psycholinguistic
research so far has focused on Germanic and Romance languages,
which are less inflecting than, for example, Slavic languages. An
exception is the study by Bordag & Pechmann (2009), which
explores declension and conjugation classes in Czech. In the
following section, this study is discussed in more detail as it
provides evidence for the cognitive reality of conjugational class
as internal feature for verb processing in Czech.
Like most Slavic languages, Czech has a pronounced
inflectional system. Syntactic functions are expressed by
inflectional suffixes on nouns, verbs and adjectives. Czech nouns
are declined according to case, number and declension class (DC);
gender has syntactic implications. Most grammarians agree upon
14 DCs, with each class denoting a specific set of inflectional
endings (see Bordag & Pechmann (2009) for an overview). Verbs
are inflected for person, number, mood, tense, voice aspect and
conjugational class (CC). There are five main classes; items
belonging to the same class share their inflectional suffixes.
Bordag & Pechmann (2009) explored the declension classes of

nouns and the conjugation classes of verbs in Czech. Both are
assumed to be lexically specified grammatical features, i.e.
internal features stored at the lemma level which have
morphological implications.
Three experiments investigated a DC/CC congruency effect
using the PWI paradigm. Participants named pictures of objects in
dative singular while they were presented with DC‐congruent or
DC‐incongruent distractors in Experiment 1, and they were asked
to name depicted actions in third person singular present tense
while they were presented with CC‐congruent or CC‐incongruent
distractors in Experiment 2. Noun stimuli were chosen from six
DCs and distractors were presented in citation form (nominative).
Verb targets and distractors came from three CCs and distractors
appeared in the infinitive. All targets were also paired with
identical and neutral (xxxxx) distractors as control conditions.
There was no interval between the presentation of the target and
the distractor (SOA = 0 ms).
As expected, incongruent distractors (DC and CC) delayed
articulation latencies and error rates were higher than in
congruent conditions. The configuration of the results resembled
the gender congruency effect. Bordag and Pechmann stipulated
generic DC/CC nodes on the lemma level which are linking
(“mediating”) elements to the inflectional endings on the word
form level. Incongruently related picture‐distractor pairs activated
two different class nodes at the same time, which strove for
activation and final selection. The competition delayed the
formulation processes. A congruently activated node, however,
bundled the whole activation and was therefore faster and more
easily selected.
Their Experiment 3, in which only DC was under investigation,
explored whether the competitive process observed resulted from
competition between abstract feature representations during the
grammatical encoding or between inflectional endings, during
phonological encoding. In the experiment, Czech participants
were now instructed to produce nouns in the genitive and
instrumental, respectively. Inflectional endings for the selected
DCs in the genitive were formally distinct; in the instrumental

they were identical. The congruency effect showing up in both
case naming conditions (genitive and instrumental) was
interpreted as competition on the abstract DC level; it could not
have its origin on the word form level because, in the
instrumental, there was overtly no mismatch and therefore no
source for any competitive process.
However, the authors interpret the effect with caution. It was a
null result (no difference between genitive and instrumental case)
and effects of grammatical and phonological encoding do not
necessarily have to be additive. It is possible that a phonological
component in the genitive singular was present, but not
statistically detectable.
With this congruency effect, the study demonstrated that DC
and CC have psychological reality and that the features are stored
generically, otherwise no competition could have been observed.
5 Tense
Tense is a grammatical category of verbs expressing the time at,
during, or over which a state or action denoted by a verb occurs.
In most languages tense is conveyed by inflectional morphemes.
This thesis focuses on tense because tense in German is
intertwined with the inflectional processes triggered by person
and number agreement between verb and subject as demonstrated
in Figure 2.
The first paragraph of this section is about the deictic character
of tense (Reichenbach, 1947; Levelt, 1989). Then the basic ideas
about linguistic (typological) placement and the formal variations
in expressing the concepts of time are discussed. The next part
provides observations about the usage (pragmatic status) of
present and past tense ‐ the two tenses under investigation.
Many theories about tense are based on the ideas of
Reichenbach (1947), a German philosopher of science. Tense is the
time an action or state happens as denoted by the verb.
Reichenbach (1947) formalised tense as a relation between S
(speaking time) and E (event time). E can also be an interval
(Comrie, 1987). These two parameters describe the three absolute
tenses:
(11) a. S = E, present [S with E]

b. S > E, past [S after E]
c. S < E, future [S before E]
Thus, in a. both the event and the report occur simultaneously.

Events in the past tense precede the moment of speaking, while
those in the future have not yet taken place.
83
84 CHAPTER 5: TENSE
In order to analyse complex, relative tenses (e.g. past perfect),

Reichenbach (1947) introduced a reference point ‐ another event or
fixed time. The reference point in (12)
(12) I called John, but he had left.
is the main clause I called John. It is distinct from the event point
(John’s leaving) and S, the moment of speaking. That elucidates
the deictic character of tense (Levelt, 1989). Tense expressions are
understood relative to the moment of speaking and deictic means
relating entities to a reference point. We can visualise tense by
locating situations on a timeline relative to the present moment or
to each other (Comrie, 1987). According to Comrie (1987),
communicating in the present tense sets “the present moment as the
deictic center”.
Closely connected with the question of the semantics of time
reference are formal factors. Not all languages have
grammaticalised tenses (Comrie, 1987; 1989). For example. the
isolating language Chinese refers to the past with adverbials.
Present and past tense in German are synthetically formed tenses
expressed by the verb (i.e. one word). The verb is composed of
several bound morphemes. Furthermore, it is necessary to
differentiate synthetically formed tense expressions from
periphrastic ones such as the past perfect, which consists of an
auxiliary and the participle form of the main verb.35 Synthetically
formed tenses may differ in the degree of tightness of the bond
between the inflectional morphemes and the verb stem. Bickel &
Nichols (2007) differentiate constructions along a scale of fusion
(isolating>concatenating>non‐linear). German tense morphology
comprises both morphemes that can be separated from their host
in a linear fashion (in the present and regular past) as well as
morphemes which defy linear segmentation from the verb stem.
The latter belong to the non‐regular past and consist of Ablaut or
Umlaut (cf. Bickel & Nichols, 2007).
35 These distinctions apply to individual constructions, like the formation of a

particular tense, rather than to whole languages.
Present tense is formally unmarked in most languages (Bybee,

1985) and thus not expressed by a separate morpheme.
Pragmatically, the present tense can also be found with generic or
future meaning. Take habitual actions like (13) as an example:
(13) I call John for lunch (at 1 o’clock).
The act of calling John must not be taken literally at the moment of
speaking. Only very rarely the time of speaking (S) and the event
time (E) occur really simultaneously. A real coincidence (see (6.a)
exists only with performative acts (cf. Comrie, 1987). A similar
divergence between formal tense marking and expressed meaning
exists for past tense as well: past forms can also be found in
conditional clauses with non‐past meaning.
According to Dudengrammatik (1998), present (52%) and past
tense (38%) are the most frequently used tenses in German and
together account for 90% of all occurrences of tensed verbs.
However, DUDEN uses written corpora, in which the past tense is
relatively frequent. In Henning’s (2000) analysis of spoken
language, the percentage of past tense immediately plunged to a
mere 15% of all sentences.
The German past tense – which formally corresponds most
closely to the English simple past – has a different pragmatic
status in German than simple past has in English. In German, and
especially in spoken language, the past tense is much less frequent
than the simple past is in English, due to the commonly used past
perfect (e.g. du hast ge‐spiel‐t [you have‐2.sg.PRES play‐PART, you
played]) referring to completed actions. Careful disentanglement
of the factors of frequency and regularity is necessary when
addressing issues of verb inflection. Consequently, assumptions
about the mechanisms governing regular vs. non‐regular
inflection in English cannot be transferred to German
immediately, without further consideration.
6 The empirical stance
6.1 Why and how regularity might be represented
The previous theoretical section was devoted to elaborate a wide‐

coverage framework of producing regular and non‐regular verbs
also reviewing research on the activation of grammatical features
during lexical access. Based on the theoretical framework and the
results from Bordag & Pechmann (2009), it is hypothesized that
regularity is mentally represented and initiates Dual Route
processing as claimed by several authors (Pinker, 1999; Clahsen,
1999; Jaeger et al., 1996). The goal of the proposed study is to
further examine the mechanism responsible for lexical selection of
verbs and to provide an initial attempt to look at the way in which
lexical representations may act as a cue for morphological
processing in one way or another.
Regularity is – like DC/CC in Czech ‐ an internal and
invariable, indispensable grammatical property with
morphological implications. Pinker & Ullman (2002) already
suggested a stored past tense feature associated with non‐regular
verbs’ lexical entries (see section 2.3.3). Three reasons support the
representation of regularity as part of the lexical information
stored in the mental lexicon connected to the appropriate lemma.
First, there are patients with neurological disorders like AW (a
patient with acquired word‐finding difficulties caused by a left‐
temporal lesion; Miozzo, 2003) showing a clear dissection of non‐
regular forms of both verbs (e.g. found) and nouns (e.g. children).
One possible explanation is that the rule‐based system is still
intact, while the non‐regular forms, however, cannot be retrieved
anymore. Then, it would be quite a coincidence that nouns as well
87
88 CHAPTER 6: THE EMPIRICAL STANCE
as verbs were affected by AW’s lesion. It is even more surprising

that while the retrieval of non‐regular word forms from the
lexicon is so severely impaired the retrieval of regular stems from
the lexicon is not affected by the damage. Although
neuropsychological association data are not the most meaningful
source of evidence, AW’s performance data might also tell that
there is a superordinate feature with two values: regular and non-
regular (cf. chapter 2.3.2.3). From theoretical considerations and
empirical data, several researchers, including Ullman et al. (1997),
Patterson et al. (2001) and Miozzo (2003), drew the following
conclusion: Patients know that a verb is non‐regular
“…and hence do not append the suffix -ed to its stem. But because
phonology remains inaccessible, they can only guess the verb form; so,
for example, they change a vowel or a consonant to produce verbs that
resemble existing non-regular verb forms. In other words, it seems that
the regular/irregular [non-regular, HT] status of the verb is
information that is stored separately from the form of the irregular verb
[non-regular, HT].” (Miozzo, 2003:124)
The article continues with a proposal similar to Pinker’s (1991):
“Perhaps the function of the information about the regular/irregular

[non‐regular, HT] status of the verb is to block the suffixation process
and to trigger the retrieval of the irregular [non‐regular, HT] form
from stored phonology.”
Second, it is not inherent to verbs’ meaning or form (i.e. infinitive,

compare geben [irregular, to give] and leben [regular, to live])
whether they follow regular or non‐regular inflection. Regularity
is an arbitrary characteristic that needs to be inherent to the lexical
entry.
Third, in addition to the theoretical considerations, Bordag and
Pechmann (2009) gave evidence for the cognitive reality of a
similar internal feature of verbs in Czech, namely conjugational
class. Understanding the regularity encoding of CC might be
helpful. CC and regularity of verbs have several things in
6.2 THE REGULARITY CONGRUENCY EFFECT 89
common. CC is also assumed to be represented at the lemma level

and to classify verbs. CC, like regularity most probably too, has to
be set before further processing. Both differ from gender as they
have morphological implications on verbs.
Summing up, the assumption of a psychologically real
regularity feature rests on the observed full loss of producing non‐
regular forms of the agrammatic patients like AW, the
arbitrariness of a verb’s regularity and the DC/CC congruency
effect (mainly CC) in Czech. The question emerges of how
regularity is represented. One possibility is the idea of generic
nodes, like the assumed representation for gender (cf. Schriefers &
Jescheniak, 1999, for grammatical gender) or grammatical class in
the most famous model of speech production (Levelt et al., 1999)
or DC/CC (Bordag & Pechmann, 2009). In this case, each lemma
would be connected to one fixed value for its regularity.
Therefore, the production of regular and non‐regular forms can
proceed very fast and is not prone to errors. This feature can be
assumed to be binary, with the two observable values being
regular and non‐regular.36.
The hypotheses in the following section are derived from the
DC/CC congruency effect in Czech, while being aware that a
priori functional and descriptive similarity does not necessarily
demand the same processing principles.
6.2 The regularity congruency effect
The experiments described in chapter 7 address the representation

and processing of regular and non‐regular verbs. The
experimental question is whether verb lemmas are stored together
with a feature that specifies their regularity status. The PWI
paradigm is used to reveal and to study the activation and
encoding of this status. The same technique already revealed
congruency effects for gender (Schriefers, 1993) and conjugational
36 It might turn out that regularity is a ternary feature if hybrid verbs are an
autonomous verb type besides regular and irregular verbs.
class in Czech (Bordag & Pechmann, 2009). With the hypotheses

established in this thesis, Experiments 1‐4 of the current study
search for a regularity congruency effect. The relations between
target and distractor were manipulated with respect to the
regularity of verbs: picture and distractor were either congruent
(both regular or non‐regular) or incongruent (they differed in
regularity). If interference occurs, it is supposed to reflect
competition for selection among regularity features of lemmas –
though this reasoning can only be analogue to Schriefers (1993)
and Bordag & Pechmann (2009). Gender and regularity are not
directly comparable because the former is a syntactic property.
Most inflecting word classes (e.g. adjectives, pronouns, numerals)
require agreement in gender (and number and case) with the head
noun. DC/CC and regularity are more closely related. CC is
comparable with regularity as both are morphological properties
of verbs.
Temporal changes in the mapping of pictures onto their
respective phonological forms may in effect provide insight into
encoding processes during lexical access. For production in
general, comparing naming latencies gathered with the PWI
paradigm, different naming conditions can be placed along a
latency scale (cf. the scale in (14)). Naming without distractors is
fastest (cf. Schriefers et al., 1990) and can serve as a baseline for
other conditions. The mere presence of a string of characters
increases articulation latencies. Even a neutral distractor like a
row of ‘x’ (xxxxx) delays naming. This condition measures the
interference of any stimulus. Identical distractors lead to an
increase in reaction time, but not as much as critical distractors
which embody the experimental manipulation. Their placement
on the scale depends on their nature.
(14) without < xxxxx < identical < critical distractors
Specifically, the most interesting relation is the one between the

critical conditions. The hypotheses are based on the CC‐
congruency effect (Bordag & Pechmann, 2009) because of the
striking resemblance between CC in Czech and regularity in
German. It is expected that naming in incongruent conditions is

slower than in congruent conditions because of a Stroop‐like
interference effect. Competition between different regularity
values should prolong reaction times to the target. That is, if
participants consistently name pictures in incongruent trials more
slowly than in congruent trials, this reflects additional effort to
activate the target including its grammatical features. Conflicting
information activated through the distractor must be suppressed.
It takes time to resolve the conflict37. Further, it is predicted that
naming is fastest in control conditions (xxxxx and identical) for
which I expect no interference due to the regularity status of the
target. Effects shown experimentally may hint at the mechanisms
of language production in an actual discourse context – here,
various other interference phenomena would have to be assumed.
In addition to the predicted regularity congruency effect, it is of
interest to look for a pure regularity effect in present tense, like the
one found for past tense naming by Seidenberg & Bruck (1990, as
cited in Seidenberg, 1992). In their study, regularity clearly
affected articulation latencies. The authors had participants inflect
regular and non‐regular verbs in past tense from written present
tense stems and measured onset latencies. Though the materials
were matched for present tense stem frequency, it took
participants significantly longer to generate non‐regular past tense
forms. On average 100 ms deviation between regular and non‐
regular verbs’ articulation latencies were reported and this result
equals Prasada et al.’s (1990). The Words and Rules Theory
(Pinker, 1999) and the theory about internally structured lexical
entries (Clahsen, 1999) assume that regular as well as non‐regular
verbs are inflected regularly in present tense. Pinker (1999) does
not deal with this issue explicitly, but the theory implies that verbs
which can be built regularly are computed regularly. Therefore,
37 Things are reverse regarding word class. Competing word class information
facilitates naming, because words of different grammatical categories do not fight
for the same syntactic slot (Pechmann & Zerbst, 2002; Pechmann & Garrett, 2004)
but words of the same grammatical category do. Regularity, however, has more
similarities with CC and therefore hypotheses are derived from Bordag &
Pechmann (2009).
only past tense non‐regular verbs are assumed to be fully listed.

Similarly, present tense regular status of verbs is implied, though
not expressed explicitly by Clahsen (1999): here, the present tense
word forms are the top nodes of the underspecified lexical entries
(cf. Figure 4, section 4.1.2). Non‐regular word forms are embodied
by sub‐nodes which are limited to past tense. Contrary to these
two theories, Janssen (1999) found signs of non‐regular processing
of Dutch verbs also in present tense. He investigated inflection
frames of verbs which were specific for verb type (regular or non‐
regular verbs) independent of the tense produced. Janssen (1999)
concluded that the inflectional frame of non‐regular verbs does
not have a specific slot for the tense suffix (e.g. –te for German
past tense) because their past tense is lexically coded, i.e.
idiosyncratic. As an effect of regularity it is expected that it takes
less time to produce regular verbs:
(15) regular < non‐regular
If verbs maintain their regularity status across tenses, the

regularity effect is hypothesised to appear also across tenses.
Hence, if this regularity effect can be replicated in present tense,
where regularity for most German verbs is not overtly visible,
verbs are likely to be stored and represented as classes, i.e.
regularity is then represented as an abstract generic feature. Those
verbs connected to the same feature are inflected alike.
Alternatively, one might argue that word forms are marked for
regularity individually, i.e. only by means of their past tense form.
Under the assumption of individually marked word forms, the
regularity effect is not assumed to arise in present tense because
regular and non‐regular verbs do not belong to the same
morphosyntactic category. In this case, regular and non‐regular
verbs are expected to be superficially regular and therefore to be
equally fast.
Intimately connected with the question of how regularity
might be represented is the question on the structure of lexical
entries of verbs. In particular, the retrieval of non‐regular word
forms might be more difficult because the lexical entry contains
several word forms from which a single one must be selected. If

the regularity effect hypothesis in the present tense proves true, an
explanation different from both the blocking mechanism (Pinker,
1999) and the assumption of internally structured lexical entries
(Clahsen, 1999) is more likely – relying more on the structure of
lexical entries – for example its complexity.
As this is, to my knowledge, the first time the present tense
production of verbs is studied, predictions for the factor TENSE
are drawn from pragmatic factors: German past tense is used less
frequently than present tense (cf. section 5) Therefore, naming of
actions in present tense is expected to proceed faster than naming
in past tense.
(16) present tense < past tense
It will be examined whether the tense of the utterance affects

regular and non‐regular verbs in a similar manner. Presumably, a
tense effect (the past tense is produced more slowly than the
present tense) occurs only for non‐regular verbs, since they utilise
different word forms in either tense and past tense word forms are
less frequent than present tense word forms. Regular verbs,
however, are assumed to employ the same stem, so the discourse
frequency of the tense should not matter.
In summary, the hypotheses tested in Experiments 1 to 5 are the
following:
(1) The articulation process is delayed by neutral distractors

(xxxxx), identical distractors, congruent and incongruent
distractors listed in order of increasing interference
(2) Regular verbs are produced faster than non‐regular verbs in
past and present tense
(3) Verb naming of non‐regular verbs in the present tense is
faster than verb naming in the past tense. Regular verbs are
produced equally fast in both tenses
However, the first hypothesis cannot be examined in Experiment 5

using the modified experimental paradigm, namely simple picture
naming without distractors.
Depending on the results and the interaction between Type of
distractor, Regularity and Tense, conclusions can be drawn about:
(1) whether verbs are stored in classes.

(2) whether regularity is represented as a generic feature.
(3) whether tense contributes to Dual Route processing.
(4) which mechanism can explain present tense performance.
(5) inflectional encoding with underspecified lexical entries.
Previous studies investigated either past tense or participle

formation of regular and non‐regular verbs. With the hypotheses
sketched, the current study aimed at the production of present
and past tense. The focus in the following empirical investigation
is on whether there is a dissociation between construction or
retrieval of tense marked verb stems and to explore whether tense
affects the production processes. The hypotheses implicate the
presupposition that German verbs behave in the same way as the
English verbs, even though German has a richer morphology. Still,
this difference is only crucial for person/number marking and
should not affect the regular/non‐regular distinction with regard
to the generation of tensed stems. In both languages, the tense‐
marked stem is assumed to be generated first and then inflected
for person and number.
7 Experiments
7.1 Experiment 1 – Present tense
To review, the main goal of Experiment 1 was to evaluate the

representation of regular and non‐regular German verbs in
present tense where regularity differences are not apparent.
However, Janssen (1999) pointed out that strong verbs of Dutch
showed signs of idiosyncrasies in present tense. Exploring present
tense allows to draw conclusions as to the way of representation
and storage (paradigmatically or individually) of verbs.
For this purpose, a variant of the picture‐word interference
paradigm was adopted. The gender congruency effect as well as
the word class and DC/CC effect have been reliably observed in
previous studies involving that paradigm in several languages
(German: Pechmann & Zerbst, 2002; Pechmann, Garrett & Zerbst,
2004; Dutch: Schriefers, 1993; Italian: Alario & Caramazza, 2002;
Schiller & Caramazza, 2003; Czech: Bordag & Pechmann, 2009).
During the presentation of the picture and the distractor, both
lemmas and the associated feature nodes become active. If picture
and distractor activate two distinct features, more information
must be gathered before the competition can be resolved that the
correct name for the picture can be selected. To assess this,
articulation latencies under both congruent and incongruent
conditions will be compared with each other as well as with
control conditions. Speech onset latencies are a dependent
measure.
The Stroop‐like task is only a tool to demonstrate cognitive
processes. Normal production processes might proceed mostly
without disruptions. The advantage of the picture‐word
95
96 CHAPTER 7: EXPERIMENTS
interference paradigm in speech production is that it involves

conceptualisation and avoids potential priming between the
presented and elicited forms. However, Experiment 5 explores the
limit of the methodology of Experiments 1‐4 and also serves as a
control experiment. It is supposed to demonstrate separable
morphological processes during morphological encoding and
hence any effects will be located at the word form level (as
opposed to the lemma level in Experiments 1‐4).
7.1.1 Methods
Participants
Thirty‐two native speakers of German participated in Experiment

1. They were all students of the University of Leipzig. All received
money for their participation.
Materials
A set of 32 black and white line drawings was selected for the
experiment. The pictures depicted the actions of intransitive
German verbs (see example Figure 6). Twelve additional pictures
served as practice items. The pictures were taken from Masterson
& Druks (1998) or were created in comparable style and
complexity. Sixteen verbs were classified as regular, the others as
non‐regular (cf. Appendix) and selected items were controlled for
word form and lemma frequency (Baayen, Piepenbrock & van
Rijn, 1993).
Each picture was paired with four different distractor words
that were either identical, neutral or ‐ with respect to regularity –
congruently or incongruently related with the item (see Figure 6).
In the identical condition, the distractor was identical with the
name of the picture (picture = “singen” [to “sing”] – distractor =
singen [to sing]). Identical distractors make congruency effects
more likely. Participants seem to pay more attention to the
distractors if some are identical (Pechmann & Schriefers,
7.1 EXPERIMENT 1 – PRESENT TENSE 97
unpublished). In the neutral condition a row of five Xs was used

as distractor (“singen” ‐ xxxxx). In the congruent condition the
verbs used as pictures and distractors were either both regular or
both non‐regular. In the incongruent condition they differed in
regularity.
Figure 6. Stimuli examples for the verb trinken (to drink).
Combined target verbs and distractors were neither semantically

nor phonologically related. Semantic relatedness was judged by
two independent native speakers. To avoid phonological
facilitation, pairings were restricted not to share initial segments,
not to rhyme or not to belong to the same ablaut pattern. These
criteria restricted the target pool enormously. Thus, it was not
possible to match the groups of regular and non‐regular verbs
exactly for length and frequency (see section 7.2.3 for a discussion
of these issues).
All distractors belonged to the response set, i.e. they were
pictures in the experiment. Previous studies demonstrated (Levelt
et al., 1999) that competition among lexical items is stronger if
distractors are possible responses. They are previously sort of
flagged in the familiarisation phase and therefore their activation
level is higher (Levelt et al., 1999). This boosts interference38 (cf. as
well Mahon et al. (2007) exploring “response related criteria”).
38 However, Caramazza & Costa (2000) did not find support for this assumption
studying semantic interference The current study will test morphological
Apparatus
The experiments took place in a dimly lit noise insulation booth at

the University of Leipzig. The stimuli were presented as black line
drawings on a light grey background of a computer screen using
ERTS (Experimental Run Time System, Beringer, 1995).
Distractor words were presented in lowercase black letters in
bold font. Response times were measured by a voicekey. The
Microphone used was a Sony MS957.
Procedure
Participants were tested individually in a quiet room. They read

written instructions requesting that they should name the
depicted action as fast and as accurately as possible. They could
ask the experimenter to clarify the whole procedure.
The first part of the experiment was designed to familiarise the
participants with the stimuli. In a first step they named each
picture in the 3rd person singular present tense in a sentence
context (Jemand singt. [Somebody is singing.]). In that phase, the
appropriate verbs were written under the picture in the infinitive,
such that participants could read the correct picture names and
memorise them. In a second step, all pictures should be named
again without the appropriate words written underneath the
images. Finally, the participants performed 12 practice trials that
were identical to the experimental situation. After the practice set,
participants received feedback if necessary (e.g., a remark on their
clicking noises).
The second part was the experiment itself. The sequence of
events in a trial was as follows: a fixation‐asterisk appeared in the
centre of the screen for 1000 ms. This indicated that the subject of
the sentence „jemand“ [someone] had to be produced. 200 ms after
the disappearance of the fixation asterisk, the picture and
distractor were shown in the centre of the screen. The depicted
interference. Just to be on the safe side the distractors were taken from the response
set.
action was uttered in the 3rd person singular present tense, thus a
correct grammatical sentence was produced by the participants.
Participants’ reaction times were measured from displaying the
picture on the screen until the first phoneme of the utterance
triggered a voice key. After triggering the voice key, the picture
and the distractor disappeared from the screen.
To refrain overlap of inflectional endings, distractors were
presented in the infinitive. Furthermore, the distractors were
presented with four stimulus onset asynchronies (SOA: ‐200, ‐100,
0, 100 ms) randomly below or above the target picture as written
words. Participants were instructed to ignore them and to simply
name the picture. Every target picture appeared once in each
condition and SOA.
The experiment consisted of four blocks – one for each SOA.
They alternated across participants. Each target appeared once in
each block, four times in the whole experiment. Target‐distractor
pairs were rotated among the four SOA. For each SOA, the targets
were presented once in the congruent condition, once in the
incongruent condition, once in the neutral condition and once in
the identical condition. The different distractor conditions in each
block were randomised across the participants.
Each block started with two warm‐up items. Those were
excluded from all analyses. The experimenter evaluated each
utterance for correctness via keyboard after each trial. The
participants did not receive this evaluation as feedback.
The experiment lasted approximately 20 minutes.
7.1.2 Results
Incorrect responses were excluded from all analyses. Reasons for

exclusion were wrong namings, hesitations during articulation,
technical problems with the measuring and voice key triggering
(3.8 % of all data). Reaction times shorter than 200 ms or longer
than 1500 ms and those outside of two standard deviations of the
subject’s mean were identified as missing values. The total of
missing values was 8.6 % of all data points randomly distributed

across the four distractor types (see Table 1).
Table 1. Percentage of missing values (Experiment 1).
Type Percentage of missing values
Naming and measurement 3.8
Cut off 4.8
Total of missing values 8.6
Table 2. Mean Response Latencies (RT, in Milliseconds), Accuracy

Proportions (PA, in %) and Absolute Errors (Err), Varied by Regularity,
SOA and Distractor (Experiment 1).
Regularity SOA Distractor
reg nreg -200 -100 0 100 ident neut incon con
RT 653 657 658 657 654 652 622 644 672 683
PA 7.6 9.6 8.1 8.3 9.3 8.7 7.0 8.7 8.4 10.3
Err 155 197 83 85 95 89 72 89 86 105
Mean naming latencies per SOA, Distractor type and Regularity of

the utterance can be seen in Table 3.
Table 3. Mean Response Latencies (RT, in Milliseconds, standard

deviations in parentheses), Varied by Regularity, SOA and Distractor
(Experiment 1).
Stimulus onset asynchrony

-200 -100 0 100 M
Distractor reg nreg reg nreg reg nreg reg nreg reg nreg
identical 616 622 621 608 620 638 633 627 622 623
(94) (83) (99) (84) (99) (96) (99) (86) (98) (88)
neutral 647 645 649 645 642 648 644 623 645 643
(99) (92) (98) (103) (103) (99) (102) (104) (100) (100)
incongruent 672 677 672 683 655 684 663 669 665 678
(103) (104) (107) (94) (108) (118) (119) (116) (109) (108)
congruent 688 690 693 694 671 676 672 683 681 686
(110) (113) (99) (110) (124) (126) (132) (136) (117) (122)
M 655 661 658 656 647 661 653 650 653 657
(105) (101) (104) (103) (110) (111) (115) (115) (109) (108)
Naming latencies were entered into a within‐subjects analysis of

variance (ANOVA) with three independent variables: SOA (four
levels), DISTRACTOR TYPE (four levels) and REGULARITY of
the uttered verb (two levels). In addition to the participant
analyses (F1), an item analysis (F2) was conducted. Only
significant F‐ratios are reported.
A main effect of DISTRACTOR TYPE was found [F1(3,93) =
67.04, MSE = 3001.31, p<.001; F2(3,45) = 66.63, MSE = 1473.33, p <
.001] but none for SOA or REGULARITY. The post hoc Scheffé test
for the main effect (diffcrit; p<.05. = 17.1) indicated that the two control
conditions (identical M=622, SD=93) and neutral (M=644, SD=100)
differed significantly from each other and from the two
experimental conditions (incongruent (M=672, SD=108) and
congruent (M=683, SD=119). However, the Scheffé test showed
both critical conditions having the same amount of influence on
naming the pictures.
Interactions were found between SOA and DISTRACTOR
TYPE [F1(9,279) = 2.40, MSE = 1986.69, p < .05; F2(9,135) =1.41,
MSE = 1264.06, p = .19] as well as SOA and REGULARITY

[F1(3,93) = 2.86, MSE = 1827.31, p < .05; F2(3,45) = 1.51, MSE =
1156.27, p = .22]. However, since they were not even marginally
significant in the item analysis, and especially in light of the
missing main effects an interpretation without further
experiments seems to be too speculative to be seriously discussed.
Analysing the second dependent variable ERROR RATES an
ANOVA with the factors DISTRACTOR TYPE, REGULARITY and
SOA was performed. A main effect for DISTRACTOR TYPE
[F1(3,93) =3.9, MSE = 4.2, p<.05; F2(3,45) =3.2, MSE = 2,5, p < .05]
was found but none for REGULARITY and SOA. The posthoc
Scheffé test did not confirm the main effect. The lack of
significance shows that the errors made are equally distributed
across conditions. None of the conditions is more error prone than
the others (see Table 4).
Table 4. Accuracy Proportions (in %, absolute numbers in parentheses),
Varied by Regularity and Distractor (Experiment 1).
Distractor
Regularity inc con id neu
reg 6.6 (34) 8.8 (45) 5.9 (30) 9.0 (46)
nreg 10.2 (52) 11.7 (60) 8.2 (42) 8.4 (43)
7.1.3 Discussion
In Experiment 1, it was tested whether regularity is represented in

form of abstract regularity nodes as assumed for gender, number
or conjugational class (Levelt et al., 1999; Bordag & Pechmann,
2009). In a picture‐distractor paradigm, participants named
pictures of actions with verbs in the 3rd pers.sg. present tense. The
analyses showed a main effect of Distractor type on naming
latencies, but neither an effect of Regularity nor interactions as
had originally been hypothesised. There was no significant
difference in accuracy. The main effect of Distractor type revealed
no difference between congruent and incongruent distractors, but
significant differences to the control distractors (identical, neutral).
7.2 EXPERIMENT 2 – PAST TENSE 103
In fact, participants did not ignore the distractors, but the

congruency of the verbs with respect to regularity did not affect
the articulation process. The data do not reflect a competition of
different feature values of the presumed abstract regularity
feature. Rather, participants seem to be only affected by the mere
appearance of a distracting (irrelevant) verb, independently of its
regularity. Actually, no conclusion for morphological processing
can be drawn based on this null effect.
While regularity is not consistently relevant for present tense
inflected verbs, it is consistently crucial information for past tense
inflection. In the past tense, verbs are coherently and uniformly
clearly identifiable as regular or non‐regular. All verbs equal in
manifesting their regularity status in past tense. To test this
assumption and the regularity congruency effect, the same
experiment was run again using past tense inflected verbs, for
which the regular/ non‐regular distinction has already been
shown in former studies (Prasada et al., 1990; Seidenberg & Bruck,
1990, as cited in Seidenberg, 1992) but without addressing its
representation.
7.2 Experiment 2 – Past tense
7.2.1 Method
Participants
Thirty‐two students of the University of Leipzig participated in

the experiment and were paid for their participation. None of
them had taken part in Experiment 1.
Materials and Procedure
Methodologically, Experiment 2 was identical to Experiment 1.

The only difference was that participants were now instructed to
produce sentences in the past tense where regularity is overtly
visible for all items, e.g. Jemand sang. [Somebody was singing.].
7.2.2 Results
Naming and measurement errors (6.2 % of the data) were

excluded using the same criteria as in the previous experiment. A
cutoff discarded extreme articulation latencies shorter than 200 ms
and longer than 1500 ms as well as values plus or minus two
standard deviations from the subject’s mean. They were handled
as missing values. The total of missing values was 6.9 %. Mean
reaction times and error rates for each of the three involved factors
are summarised in Table 6.
Type Percentage of missing values

Naming and measurement 6.2
Cut off 0.7
Total of missing values 6.9

Regularity SOA Distractor
reg nreg -200 -100 0 100 ident neut incon con
RT 705 734 721 724 720 714 681 716 746 738
PA 6.6 7.1 6.7 5.9 7.8 7.1 4.2 6.4 7.6 9.3
Err 136 146 69 60 80 73 43 166 78 95
Table 7 summarises the full picture of results of Experiment 2.

Table 7. Mean Response Latencies (RT, in Milliseconds, standard deviations in

parentheses), Varied by Regularity, SOA and Distractor (Experiment 2).

-200 -100 0 100 M
identical 647 689 657 685 689 700 662 713 664 697
(128) (145) (123) (118) (159) (138) (123) (133) (135) (134)
neutral 709 748 711 734 688 715 695 723 701 730
(146) (140) (137) (137) (135) (118) (141) (142) (140) (135)
incongruent 724 769 745 766 719 791 725 730 728 764
(139) (134) (146) (137) (139) (167) (159) (146) (146) (148)
congruent 729 753 743 762 720 741 712 744 726 750
(135) (139) (141) (120) (131) (135) (139) (145) (137) (135)
M 702 739 713 735 704 736 699 727 705 734
(141) (143) (141) (132) (142) (144) (143) (142) (142) (140)
For Experiment 2, a three factorial ANOVA was performed again,

both by participants (F1) and items (F2). The results revealed
significant main effects for DISTRACTOR TYPE [F1(3,93) = 57.41,
MSE = 4082.20, p < .001; F2(3,45) = 41.92, MSE = 2799.42, p < .001]
and REGULARITY [F1(1,31) = 34.37, MSE = 6407.55, p < .001;
F2(1,15)=5.34, MSE = 20689.16, p<.05]. A post hoc Scheffé test for
DISTRACTOR TYPE (diffcrit; p<.05. = 16.0) showed that the control
conditions (identical and neutral) differ from the experimental
conditions (congruent and incongruent), but no effect was found
comparing the two critical conditions with each other. Although
the interaction SOA by DISTRACTOR TYPE F1(9,279) = 2.65, MSE
= 4111.56, p < .01; F2(9,135) = 2.40, MSE = 2333.63, p < .05 reached
significance, no significant differences could be observed
comparing the critical conditions by a post hoc Scheffé test (diffcrit;
p<.05.=66.5). The interaction stems from the longer naming latencies
under the critical condition at SOA ‐200 and ‐100 compared to the
control conditions, which is not of exorbitant interest. Interactions
with the factor REGULARITY were not significant.
Table 8 shows error rates by type of distractor and regularity of

the utterance (regular or non‐regular).
Varied by Regularity and Distractor (Experiment 2).
Distractor
Regularity inc con id neu
reg 6.1 (31) 9.4 (48) 5.1 (26) 6.1 (31)
nreg 9.2 (47) 9.2 (47) 3.3 (17) 6.8 (35)
The analysis of errors was also significant. An ANOVA with the

Distractor type (four levels) and Regularity (two levels) as
variables revealed a significant main effect in the number of errors
for Distractor type: F1(3, 93) = 7.09, p < 0.001, F2(3, 45) = 6.0, p <
0.01. A paired‐samples t‐test comparing only the congruent and
incongruent conditions showed that this effect was not significant:
t1(126) = ‐1.29, p =.197, t2(62) =‐1.01, p =.317.
Mean articulation latencies under neutral condition were
compared between both tenses. Experiment 1 yielded a present
tense mean of 644 ms and Experiment 2 a past tense mean of 715
ms. An independent two‐tailed t‐test39 of articulation latencies
between present and past tense under neutral condition showed
significance, t1(1747) = ‐12.86, p < .001.
7.2.3 Discussion
The picture‐word interference paradigm is used in speech

production research to show and elicit interference or facilitation
effects ‐ depending on the feature studied. It is assumed that
target and distractor both compete for selection and activate their
grammatical features prior to the final selection of the target.
Experiment 2 sought to elicit a regularity congruency effect due to
competing abstract regularity nodes in verb naming in past tense.
39 The results are reported despite the fact that the Levene’s test for homogeneity
was significant, which means that the assumption of homogeneity of variances is
violated.
The results indicate that critical distractors (both incongruent

and congruent) increase the difficulty to access the correct lexical
item. Hence, compared to control conditions the critical distractors
are indeed doing their job. The regularity of the distractor,
however, does not seem to affect the production, even though
being obvious in the past tense. Any word not related in either
form or meaning superimposed to the picture would have had the
same impact. What we see is an unspecific distraction effect not
applicable to the regularity of the verbs. Furthermore, the
interaction SOA by Distractor type gives the additional
information that distractors affect the target at the two earliest
SOAs, namely ‐200 and ‐100 ms.
Although the current results do not reveal the predicted
congruency effect by Distractor type, they demonstrate a strong
Regularity main effect as was also expected. It takes participants
longer to produce non‐regular verb forms than regular ones
(Prasada et al., 1990; Seidenberg & Bruck, 1990, as cited in
Seidenberg, 1992). This boosts the assumption that regular and
non‐regular morphology involve separate morphological
processes. Latency differences between regular and non‐regular
verbs are classically explained with blocking which postulates a
specific advantage in producing regular verbs (Pinker, 1999).
Using an application of the Dual Route Theory, the Dual Route
Mechanism must be triggered as argued in the introduction ‐ at
least the blocking procedure itself. The experiments do not give
evidence for such a “regularity feature” as of yet because the
critical conditions do not reveal an interference effect.
Up to this point the results do not form a unified picture,
including some contradicting elements. Since the results of
Experiment 1 do not look like regularity is influencing the
production latencies in the present tense and the past tense results
of Experiment 2, however, do point to an influence of regularity,
the factor tense seems worthy of further exploration. A
comparison of naming latencies under neutral conditions in
Experiment 1 and 2 (taking into account that the participants of
the experiments were not identical, thus skewing the reliability of
the findings) showed that it takes participants longer to name
actions in the past tense than in the present tense. This could well
be because the present tense is much more frequent in discourse
and spoken language. These findings, as well as the conclusion,
are in need of empirical support and precise evidence.
Experiments 1 and 2 tested independent samples of subjects and
should be replicated in Experiment 3 with dependent samples.
Therefore, Experiments 1 and 2 were combined into Experiment 3,
in which participants were presented with both tenses, making
tense a fourth factor. The interaction of regularity by tense
(probably by SOA as well) is of most interest.
7.3 Experiment 3 – Present and past tense
The regularity effect in the past tense (cf. Experiment 2), compared
to the missing regularity effect in the present tense (cf. Experiment
1), suggests that tense is an influencing factor for verb production.
Indeed, the possibility of regularity effects modulated by tense
seems probable. While the present tense does not necessarily have
to activate regularity, the past tense should do so. To the extent
that regularity is a property of individual forms, and to the extent
that both picture and distractor activate their feature during word
production, increased (morphological) interference in the past
tense for items that are paired with incongruent distractors is
hypothesised compared to those with congruent distractors and
compared to all present tense conditions. More pronounced
interference effects in the former than in the latter condition
indicate competition during lexical selection and may be
attributed to the differing regularity status of the interfering
stimuli.
7.3 EXPERIMENT 3 – PRESENT AND PAST TENSE 109
7.3.1 Method
Participants
Thirty‐two students of the University of Leipzig were recruited to

participate in the experiment. None of them had taken part in
Experiments 1 or 2.
The same set of 32 black and white line drawings from

Experiment 1 was used in Experiment 3. Experiments 1 and 2 in
their original form, i.e. completely, were combined to create
Experiment 3. The sequence of tense was counterbalanced across
participants. 50 percent were asked to name pictures first in the
present tense (Experiment 1) and afterwards the past tense
(Experiment 2), the remaining participants were asked to name
them the other way around (Experiment 1 after Experiment 2).
The second and third training phase, as described in Experiment
1, were included once more before the change of tense to ensure
that participants correctly named the pictures in the appropriate
(new) tense.
Stimuli presentation followed the procedures of Experiments 1
and 2.
7.3.2 Results
Again, naming and measurement errors were excluded, as well as

reaction times shorter than 200 ms or longer than 1500 ms and
latencies deviating more than two standard deviations from
subjects mean. 4.7 % of all data were excluded as outliers using
the same criteria as described in Experiments 1 and 2. Errors
identified as voicekey errors and wrong namings made up 4.5 %
of all data, in sum 9.2% of all data were identified as errors
leading to missing data. The percentages of errors separated for
each tense are given in Table 9.
Percentage of missing values

Type present past total
Naming and measurement 3.1 5.9 4.5
Cut off 3.4 6.0 4.7
Total of missing values 6.5 11.9 9.2

Regularity SOA Distractor Tense
reg nreg -200 -100 0 100 ident neut incon con pres past
RT 650 676 661 670 665 654 626 655 680 689 653 673
PA 6.9 11.5 8.9 9.7 9.2 8.9 7.7 8.3 10.7 10.0 6.5 11.8
Err 281 472 183 199 189 182 158 170 220 205 268 485
Table 11 summarises the results of Experiment 3.


deviations in parentheses), Varied by Regularity, SOA, Distractor and
Tense (Experiment 3).
-200 -100 0 100 M

PAST TENSE
identical 609 638 604 645 628 667 629 661 617 653
(91) (82) (93) (94) (90) (91) (79) (84) (89) (88)
neutral 656 679 654 706 649 699 637 674 701 730
(99) (96) (92) (102) (102) (93) (90) (91) (140) (135)
incongruent 658 714 681 715 669 723 665 707 649 689
(94) (99) (97) (90) (103) (125) (108) (114) (96) (96)
congruent 692 730 703 727 679 698 667 690 685 712
(116) (116) (104) (103) (106) (105) (113) (116) (111) (110)
M 653 689 659 698 656 696 649 682 654 692
(104) (105) (103) (102) (102) (105) (99) (102) (102) (104)
PRESENT
identical 598 617 607 629 617 630 607 631 607 627
(93) (95) (92) (97) (95) (97) (77) (94) (89) (95)
neutral 640 644 646 644 627 652 636 651 637 648
(102) (89) (94) (95) (89) (96) (89) (102) (94) (95)
incongruent 651 694 670 693 662 685 650 654 658 682
(94) (100) (101) (99) (106) (104) (108) (98) (103) (101)
congruent 676 688 691 706 670 692 664 657 675 686
(105) (118) (103) (110) (104) (110) (120) (101) (108) (111)
M 641 661 653 668 644 665 639 648 644 660
(102) (106) (102) (105) (101) (105) (102) (99) (102) (104)
An ANOVA was performed using a within‐subjects design with

four independent variables: DISTRACTOR TYPE (four levels:
identical, neutral, congruent and incongruent), REGULARITY of
the verbs (two levels: regular or non‐regular), TENSE (two levels:
present or past) and SOA (‐200, ‐100, 0, +100 ms). Dependent
variables consisted of naming latencies and percent accuracy.
The ANOVA yielded significance for all four factors:

DISTRACTOR TYPE [F1(3,93) = 92.73, MSE = 4481.68, p < .001;
F2(3,45) = 144.46, MSE = 1473.72, p < .001], REGULARITY [F1(1,31)
= 97.04, MSE = 3589.48, p < .001; F2(1,15) = 6.61, MSE = 26770.56, p
< .05], TENSE [F1(1,31) = 22.20, MSE = 10138.86, p < .001; F2(1,15) =
80.71, MSE = 1463.58, p < .001] and SOA [F1 (3,93) = 3.82, MSE =
4599.40, p < .05; F2(3,45) = 8.04, MSE = 1256.32, p < .001].
There were interactions between the following factors: TENSE
x REGULARITY [F1(1,31) = 20.21, MSE = 2673.41, p < .001; F2(1,15)
= 12.48, MSE = 2133.93 p < .01], SOA x DISTRACTOR TYPE
[F1(9,279) = 4.45, MSE = 2746.66, p<.001; F2(9,135) = 2.81, MSE =
2320.08, p < .01] and SOA x DISTRACTOR TYPE x TENSE
[F1(9,279) = 2.16, MSE = 1444.31, p<.05; F2(9,135) = 2.02, MSE =
934.77, p < .05].
Post hoc analyses of the main effects showed a significant
difference between control (identical and neutral) and
experimental (congruent and incongruent) distractors, but not
between congruent and incongruent distractors (Scheffé test diffcrit;
p<.05. = 12.0).
Table 12. Mean Response Latencies (RT, in Milliseconds), Varied by

Distractor and Response Latency difference between control and
experimental stimuli (Experiment 3).
Distractor type
control experimental
identical neutral incongruent congruent
RT 626 655 680 689
M 641 685
Effect +44
While the SOAs ‐100 ms and +100 ms showed a marked

difference, the same could not be said for the ‐200 ms and 0 ms
SOAs (Scheffé test diffcrit; p<.05. = 8.6). See Table 10 above.
The Scheffé‐test for the TENSE x REGULARITY interaction
(diffcrit; p<.05. = 26.4) revealed a regularity effect in the past tense:
non‐regular verbs were produced more slowly than regular ones
in the past tense. In the present tense they were produced with
equal speed (see Table 13).
Regularity and Tense (Experiment 3).
Regularity
Tense reg nreg
past 654 692
present 644 660
The interaction DISTRACTOR TYPE by SOA (Table 14) manifests

in identical distractors being faster than all other distractors for
SOA ‐200 and ‐100. Unfortunately, the interaction does not
originate from the reaction times of the critical congruent and
incongruent distractors. A post hoc Scheffé test for the triple‐
interaction SOA x DISTRACTOR TYPE x TENSE (Table 15)
specifies this finding for the past tense SOA ‐100 and ‐200 only:
diffcrit; p<.05. = 39.7.
Distractor and SOA (Experiment 3).
SOA
Distractor -200 -100 0 +100
identical 615 621 635 631
neutral 654 662 655 649
incongruent 679 689 683 668
congruent 696 706 684 669

Distractor, Tense and SOA (Experiment 3).
SOA
PAST
Distractor -200 -100 0 +100
identical 623 624 647 644
neutral 667 679 673 654
congruent 711 715 688 678
PRESENT
Distractor -200 -100 0 +100
identical 608 618 624 619
neutral 642 645 639 643
congruent 682 698 681 661
Error analysis
Inspection of the Accuracy proportions allows for the

presumption of a tense effect (cf. Table 16). An Anova was
computed with the factors TENSE, DISTRACTOR TYPE and
REGULARITY.
Varied by Regularity, Tense and Distractor (Experiment 3).
Distractor
Tense Regularity inc con id neu
past reg 9.8 (50) 11.7 (60) 6.8 (35) 6.1 (31)
nreg 18.4 (94) 14.6 (75) 13.3 (68) 14.1 (72)
present reg 5.5 (28) 6.1 (31) 3.9 (20) 5.1 (26)
nreg 9.4 (48) 7.6 (39) 6.8 (35) 8.0 (41)
The error analysis of all missing values confirmed significant main

effects for the three factors: DISTRACTOR TYPE [F1(3,93) = 7.09, p
< .001; F2(3,15) = 5.06, p < .01], REGULARITY [F1(1,31) = 47.7, p <
.001; F2(1,15) = 5.6, p < .05], TENSE [F1(1,31) = 28.9, p < .001;
F2(1,15) = 24.7, p < .001]. A paired‐samples t‐test comparing only

the congruent and incongruent conditions revealed that the type
of critical distractor does not have an effect: t1(254) = 0.65, p =
0.515, t2(126) = 0.43, p = 0.667. The analyses of errors revealed that
the number of errors was statistically the same in the congruent
and incongruent conditions.
7.3.3 Discussion
Experiment 3 strongly resembled the previously described

Experiments 1 and 2 in question and method. However,
Experiment 3 addressed the influences of tense on the
hypothesised abstract regularity node of verbs more closely. Tense
was established as within‐subjects factor (besides Regularity, SOA
and Distractor type).
Although the data of the PWI experiment revealed four
significant main effects, naming latencies did not differ regarding
the type of critical distractor as in the previous experiments. In
fact, participants did process the distractors, as shown by their
delayed responses in the experimental conditions compared to the
control conditions as before. But the distractors did not induce
interference of two simultaneously activated abstract feature
nodes in the incongruent condition. There was neither an
interaction of Distractor type with Tense nor Regularity as had
originally been hypothesised. It seems unlikely, then, that
Distractor type is affecting verb production (especially
morphological processing). This finding is in sharp contrast to the
Bordag & Pechmann (2009) study, demonstrating the
psychological reality of conjugation and declension classes and
their encoding in Czech. The congruency effect observed in that
study was interpreted as a result of the activation of generic
grammatical features.
In this experiment we thus replicated the Tense effect, now in a
within subjects design. The Tense main effect shows that the
production of past tense forms results in longer articulation
latencies than present tense forms. One possible explanation is the
less frequent occurrence of past tense forms in the language (see
section 5). The Tense main effect thus could be interpreted as a

pure frequency effect. Frequently accessed words have lower
activation thresholds and memory traces and therefore are
retrieved faster for production. This is in accordance with the Dual
Route Mechanism. Considering the Tense by Regularity
interaction, the factor Tense only plays a role for the production of
non‐regular verbs as proposed because they are assumed to be
stored. Non‐regular verbs are produced more slowly than regular
verbs in the past tense. Regular verbs are not affected by Tense
because their word forms can be built online and are not
necessarily stored in memory. Generating regularly built word
forms does not depend on memory processes (except retrieving
the stem) and activation thresholds. This is supported by the
observation that present tense naming yielded fewer errors than
naming in the past tense and non‐regular verbs caused more
errors in the past tense than in the present tense, but both failed to
reach significance. To sum up, in German, careful disentangling of
the frequency and regularity factors is necessary when addressing
this issue further.
The Tense x Regularity x SOA interaction demonstrates the
time course as found in previous experiments in the same time
window, i.e. the interaction is a replication and, hence,
methodological changes in the design of the experiment (the
addition of the fourth factor Tense) are not the reason for the
observed null effect.
Combining the missing Distractor type effect with the
Regularity effect in the past tense, it is indeed plausible that no
Distractor type effect arises in the present tense. Regular and non‐
regular processing are equal because it is not necessary to
distinguish for regularity40. But the regularity effect in the past
tense gives reason to believe that regularity is cognitively
distinctive in the past tense. It remains an unanswered question,
then why the expected interaction Regularity x Tense x Distractor
type rests insignificant. It seems to be an amazing, paradoxical
situation that in spite of the pronounced Regularity effect, naming
40 The majority of verbs follow a completely regular conjugation in present tense.

7.4 NON‐REGULAR VERBS REVISITED 117
under different distractor conditions does not lead to interference:

the Regularity main effect and the interaction Regularity x Tense
point to different processing of regular and non‐regular verbs,
which consequentially should cause modulation of interference. If
that is not the case, it is necessary to modify the hypothesis or,
respectively, accept the alternative hypothesis (H0).
So far, we do not have any empirical evidence for an abstract
representation of regularity. This is not easy to accept because
Bordag & Pechmann got very reliable congruency effects with the
same paradigm for conjugational class in Czech which is like
regularity in German arbitrary grammatical information with
morphological implications. However, the results must be
interpreted cautiously, since they may reflect errors by accident.
On the other hand, consistent evidence was virtually impossible
because of the inconsistency among items with respect to the
visibility of their regularity status in the third person present
tense: some of the non‐regular verbs involved umlaut (e.g. er gräbt
[he is digging]) whereas others employed alleged regular stems
(e.g. er trinkt [he drinking]). This difference could have been a
determining factor increasing the variances in the data and
therefore also contributing to the pattern of results.
Taken together, it seems appropriate to conduct one more
experiment, this time further controlling the material for some
additional factors. The major change was to redefine and divide
non‐regular verbs and to include a third type of German verbs,
namely hybrid verbs, as explained in the introduction.
7.4 Non-regular verbs revisited
Because of the theoretical importance of the regularity congruency

effect, Experiment 4 again tested for the Distractor type and
Regularity effect on target naming latencies, using different
materials. Linguistically, the classification of verbs in regular and
non‐regular verbs rests on verbs’ past and/or participle formation.
If a particular verb employs a non‐regular stem in either one or
both forms, the verb is said to be non‐regular. The question is
whether the cognitive representation corresponds to the linguistic

criteria.
It should be pointed out again that some non‐regular verbs in
Experiments 1‐3 were object to umlaut processes, i.e. non‐
regularities, even in the third person singular present tense (e.g.
graben ‐ er gräbt [to dig – he is digging]). In the material, present
tense stimuli therefore show great variability and it cannot be
excluded that this is mirrored by great uncertainty in the data.
However, present tense performance is quite interesting to
examine the representation of verbs in the mental lexicon. To
clarify the question clearly, more fine‐grained distinctions of verbs
in the experimental stimuli are necessary, such as a ‘semi‐regular’
category.
Coming back to the definitions given in the introduction, the
German verb system is set up by three basic paradigms. The non‐
regular paradigm is comprised of hybrid and irregular verbs. This
is a stroke of luck because it enables us to explore whether either
verb forms are separately, individually stored and marked as
regular or non‐regular or whether verbs belong to inflectional
paradigms which are as a whole regular or non‐regular. Hybrid
verbs turned out to account for this dissociation. If they were
produced in present tense by applying a rule like assumed for
regular verbs both should behave alike. In other words, the
question in Experiment 4 is whether hybrid verbs and irregular
verbs are processed similarly or if morphological processes
dissociate within hybrid verbs, i.e. their present (regular) and past
tense (non‐regular) production. For paradigmatic representation
of verbs a tense specific pattern to our results is expected.
7.5 Experiment 4
Experiment 4 aims to clear up the Distractor type null effect and

resolve the question of whether regularity is a property of whole
verb paradigms, not of individual verb forms.
If it was the regularity of particular word forms instead of
paradigms that caused the pattern of the obtained regularity effect
7.5 EXPERIMENT 4 119
for non‐regular verbs in past tense only, the same pattern will be
found in the replication. For the irregular verb type it is expected
that both past and present tense forms are irregular and therefore
do not dissociate in reaction times. It is a priori not obvious how
hybrid verbs’ naming latencies will depend on the involved
factors. If past and present tense articulation latencies are
separable one can conclude that verb forms are not stored
paradigmatically but individually. If present and past of hybrid
verbs do not differ, the whole paradigm is marked as hybrid (or
irregular, that relation has to be checked on the basis of the data).
It is important to note another aspect of the new hypothesis:
regular verbs’ present and past tenses and hybrid verbs’ present
tenses need not necessarily fall into the same group, although all
word forms seem to be regular. The reason is that regular verbs
use the same stem for both tempora and therefore are more
frequent than hybrid verbs’ present tense stems. The latter are
expected to elicit longer reaction times than the former.
For this purpose the stimulus material was revised and
adapted. An experiment was designed that contained all three
types of verbs in three accurately matched groups.
7.5.1 Methods
Participants
Eighteen students of the University of Leipzig were recruited for

Experiment 4. They were paid for their participation.
To reduce variance in the data, a new and proper set of verbs was
built up. In Experiment 4, irregular and hybrid verbs were
considered as different verb types for the first time. Until now,
they were paid no attention in the current study, as well as in
former studies. The three groups were equated in terms of
frequency, length, initial phoneme (Pechmann, Reetz & Zerbst,
1989), ablaut patterns and transitivity. Verbs containing
allomorphy (ə‐epenthesis bluten, ich blute, ich blutete versus none in

lachen, ich lache, ich lachte) were strictly excluded to avoid ə‐
epenthesis to affect reaction times as it is probably an additional
process. Materials were counterbalanced so that each item
appeared in each condition. Nine intransitive German verbs were
chosen for each group. Actions were depicted in black and white
line drawings. Some were taken from Masterson & Druks (1998),
but several were designed for this purpose in the same style and
comparable complexity.
Word‐distractor pairs were neither semantically nor
phonologically related and they were incongruent in respect to
ablaut patterns (Wiese, in prep.) to keep that factor constant.
Distractors belonged to the response set, as before.
In addition to the new set of verbs, a second step to keep the
new data clean was a reduction of item repetitions. Morphological
processes are sensitive to repetition (Zwitserlood et al., 2002)41.
Frequently and recently used word forms are assumed to be
stored in short term memory and thus easier to name on
subsequent trials. If this were the case, there would be no
morphological processes to study anymore after repeated
occurrence of the stimuli. Further, with multiple iterations, the
effect of the distractors might change across blocks. Hence, SOA
+100 was skipped, as it did not show any effects in Experiments
1‐342 and decided to leave out the neutral (xxxxx) control
condition. The identical condition was not removed, because
participants seem to recognise that distractors are sometimes
helpful and therefore read them (Pechmann & Schriefers,
unpublished). Additionally, the experiment was split in two
identical halves (but with different randomisations). Thus, the
design at least permitted to examine whether participants’
41 The authors reported that first occurrence of the stimuli were responded to
slower than second or third occurrences which did not differ from each other.
42 Another possible explanation for the insignificance is suggested by the results of
a study by La Heij and van den Hof (1995). They discovered that the semantic
interference effect size drops with (a) small target sets and (b) repeated
presentations of targets.
behaviour changed with repeated occurrences of word distractor

pairs.
The presentation of trials proceeded exactly as in Experiments
1‐3. Dependent variables were naming latencies and error rates.
The primary comparison was between the congruent and
incongruent condition like in the previous experiments.
7.5.2 Results
Voice key errors, wrong responses, hesitations and time outs were
removed from the data. The data were corrected also for outliers
which were reaction times shorter than 200 ms or longer than 1500
ms and values that exceeded two standard deviations from the
subjects mean. All in all 9.4% of all data were excluded. The cut off
was effective for 5.4 % of all data.

Cut off 5.4 5.3 5.4
Table 18 gives a summary of the mean reaction times for the four
factors in Experiment 4.

Regularity SOA Distractor Tense
Reg hyb irr -200 -100 0 ident incon con pres past
RT 635 697 707 672 686 677 644 696 694 673 684
PA 3.2 11.6 13.4 8.0 9.3 11.0 9.1 9.1 10.0 9.7 9.1
Err 63 225 260 155 180 213 177 176 195 282 266
The participants can be seen to be faster in the present tense

(M=673, SD=106) than in the past tense (M=684, SD=104) and to be
faster in the identical control condition (M=644, SD=103) than in
both experimental conditions (incongruent M=696, SD=106 and
congruent M=694, SD=100, see Table 18). Reaction times for
regular verbs (635 ms) are faster than for irregular (707 ms) and
hybrid verbs (697 ms). This visual inspection is confirmed by the
statistical analyses, which included the four factors DISTRACTOR
TYPE (three levels: identical, congruent and incongruent),
REGULARITY of the utterance (three levels: regular, irregular,
hybrid), TENSE (two levels: present or past) and SOA (three
levels: ‐200, ‐100, 0 ms).
Table 19 states the results precisely.

deviations in parentheses), Varied by Regularity, SOA, Distractor and
Tense (Experiment 4).
-200 -100 0 M
PAST
Distractor reg irr hyb reg irr hyb reg irr hyb reg irr hyb
identical 597 650 640 620 677 663 612 706 672 610 676 658
(90) (114) (100) (95) (107) (112) (85) (107) (99) (91) (111) (104)
incongruent 649 727 725 666 754 728 654 723 729 656 735 727
(82) (106) (104) (82) (106) (102) (110) (109) (115) (92) (107) (107)
congruent 642 723 722 662 744 723 651 723 722 652 730 722
(72) (92) (108) (81) (104) (96) (97) (122) (110) (84) (106) (104)
M 629 700 696 649 725 704 639 717 708 639 714 703
(85) (110) (111) (88) (110) (107) (99) (113) (111) (91) (111) (109)
PRESENT
identical 603 647 653 605 665 662 601 681 665 603 665 660
(84) (112) (121) (100) (114) (109) (90) (103) (106) (91) (110) (112)
incongruent 639 722 694 660 717 722 631 717 694 643 709 703
(91) (120) (123) (87) (100) (109) (103) (107) (111) (94) (119) (115)
congruent 655 715 705 652 730 720 626 699 707 644 714 710
(92) (111) (91) (82) (103) (107) (86) (120) (117) (87) (112) (105)
M 632 695 684 639 703 700 619 699 688 630 699 691
(91) (119) (115) (93) (109) (112) (94) (111) (112) (93) (113) (113)
A subject and items ANOVA were performed on naming latencies

with DISTRACTOR TYPE, REGULARITY, TENSE and SOA as
between subjects factors. The ANOVA yielded significance for all
four factors: DISTRACTOR TYPE [F1(2,34) = 67.62, MSE = 4085.17,
p < .001; F2(2,16) = 110.20, MSE = 1275.68, p < .001], REGULARITY
[F1(2,34) = 162.39, MSE = 3037.14, p < .001; F2(2,16) = 13.87, MSE =
18869.99, p < .001], TENSE [F1(1,17) = 5.03, MSE = 5603.75, p < .05;
F2 (1,8) = 18.37, MSE = 991.41, p <.01] and SOA [F1(2,34) = 6.90,
MSE = 2107.60. p < .01; F2(2,16) = 9.46,MSE = 902.94, p <.01].
The post hoc Scheffé test shows for SOA (diffcrit; p<.05. = 9.3) that
the reaction times for distractors appearing at SOA ‐100 ms differ
from those appearing at ‐200 and 0 ms. For DISTRACTOR TYPE
(diffcrit; p<.05. = 12.9), it revealed a dissociation between the identical
control condition and the experimental conditions (congruent and
incongruent, see Table 20).
Distractor and Response Latency difference between control and
experimental stimuli (Experiment 4).
Distractor type
control experimental
identical incongruent congruent
RT 644 696 694
M 644 695
Effect +51
The REGULARITY main effect stems from a significant difference

between regular verbs on the one and hybrid and irregular verbs
on the other side (diffcrit; p<.05. = 11.2, Table 18).
According to the ANOVA there was an interaction between the
factors SOA and TENSE [F1(2,34) = 3.55, MSE = 1700.79, p < .05;
F2(2,16) = 4.72, MSE = 457.91, p < .05]. The post hoc Scheffé test
was significant in neither F1 nor F2 (diffcrit; p<.05. = 35.4). Resolving
the interaction SOA x DISTRACTOR TYPE [F1(4,68) = 6.68, MSE =
1352.09, p < .001; F2(4,32) = 3.81, MSE = 1205.61, p < .05] with a
post hoc Scheffé test (diffcrit; p<.05. = 39.0) reveals that the control
condition differs from the experimental conditions at SOA ‐100
and ‐200 but not at SOA 0, when picture and distractor emerge
simultaneously. This finding contains information about the time
course of processing of the distractors and is a replication of the
previous experiments. Triple interactions were not significant
either. There was also no interaction between the four factors
Regularity, SOA, Tense and Distractor type.
Analyzing the first and second half separately to detect a

potential repetition effect gave the same results as the omnibus
analysis.
A curious finding is the observation that the naming latencies
of hybrid and irregular verbs did not differ from each other in
both tenses. Not only is the interaction insignificant, but even
numerical latency difference is extremely small (Table 21).
Table 21. Mean Response Latencies (RT, in Milliseconds, standard devi‐
ations in parentheses), Varied by Regularity and Tense (Experiment 4).
Regularity
Tense reg irr hyb M
639 714 703 685
past
(91) (111) (109) (104)
630 699 691 673
present
(93) (113) (113) (106)
635 707 697 679
M
(92) (112) (111) (105)
Hybrid verbs’ past and present tense production, as well as their

mean reaction times, are much closer to the latencies of irregular
verbs than they are to regular ones. As this point aroused interest,
it will be followed up on in Experiment 5, as it was not the focus
of the picture‐word‐interference experiment.
7.5.3 Discussion
Experiment 4 was designed to test for the mental representation of

regularity with a PWI paradigm. The objection raised against
Experiment 3 was met in Experiment 4. Hybrid verbs were
introduced as a third verb type to reduce variances in the data
through more carefully controlled material.
The results of Experiment 4 point to an influence of the
Distractor type, but they do not substantiate the predictions of
node‐like representation of regularity. The crucial manipulation of
the regularity relation between the target and the distractor again
leads to no result. Naming in the control condition is significantly

faster than in the experimental conditions. Verbs seem to confuse
participants, but not by their regularity. The amount of this
unspecific interference effect is exactly the same for both
experimental conditions. Hence, concerning regularity, there is
not enough evidence to decide whether verbs are stored with a
generic regularity feature. The null effects in the previous
experiments were not due to the material. Carefully checked
stimuli in Experiment 4 yielded the same results.
The results obtained in Experiment 4 are important for several
reasons. First, they demonstrate that the congruency effect from
the first experiments is robustly obscure. Second, the Regularity
effect emerged again as highly significant as well as the Tense
effect. Only the interaction between both factors revealed in
Experiment 3 could not be replicated, which is surprising.
Studying Table 18, probably the most striking aspect is that the
reaction time differences between the present and past tense
flattened. If, as presumed in the discussion of Experiment 3, Tense
is crucial for the activation and encoding of regularity in language
production, the base for the interaction is missing here.
The Experiments 1‐4 show that Regularity is involved in the
production of verbs but its representation as supposed as a
generic feature node cannot be demonstrated. It must be
represented differently as the expected congruency effect like for
encoding of gender or declension and conjugation classes was not
found. Actually, when dealing with null effects, it could be
objected that if the hypotheses are correct, the feature does indeed
exist, but the employed experimental paradigm was not sensitive
enough to reveal it. The method will therefore be evaluated in the
next section.
7.6 Discussion of Experiments 1-4
The Experiments 1‐4 did not reveal the expected regularity

congruency effect. If the hypothesis about the psychological
reality of regularity is correct, the absence of the congruency effect
7.6 DISCUSSION OF EXPERIMENTS 1‐4 127
has two main consequences: On the one hand, it casts doubts on

the adequacy of the picture‐word interference paradigm for the
study of the linguistic processes involved in verb production
although the paradigm had successfully been employed by
Roelofs (1993) and Vigliocco and colleagues (Vigliocco, Vinson,
Damian & Levelt, 2002; Vigliocco, Vinson, Lewis & Garrett, 2004;
Vigliocco, Vinson & Siri, 2005) studying verb retrieval. In the
studies mentioned, semantic interference effects similar to effects
obtained by studies concerning nouns were observed in verb
naming. On the other hand, the absence of the congruency effect
highlights the fact that the encoding of grammatical features is
probably so tightly intertwined with morphological processing
that studying one (the representation of the regularity feature)
ignoring the other may in fact prove to be impossible. The
following reasoning proceeds along these two lines.
7.6.1 Critical evaluation of the picture-word interference

paradigm
The first and most salient interference measurement was the

Stroop task (Stroop, 1935). Over the years, the picture‐word
interference paradigm was developed as an instance of the Stroop
task. In the Stroop task, the participant has to name the colour a
word is printed in, while the word itself denotes an incongruent
colour (e.g. reading the word blue written in green letters, one has
to say “green”). This results in interference and inhibitory effects
reflected by increased reaction times and higher error rates. While
one is trying to say the word’s colour (“green”), blue automatically
becomes activated through reading blue. Recognizing known
written words is a very fast and unconscious process. Automatic
actions are very hard to suppress, something that can only be
done with cognitive effort. Reading is such an action. Therefore, in
the example above, “green” and blue compete for spelling out, i.e.,
the attentional reading process is disrupted by the automatic
reading.
The picture‐word interference (PWI) paradigm follows the
same principles as the Stroop task. In PWI tasks, participants are
presented with a picture (target) and a written word (distractor).

They are instructed to name the picture, ignoring the word.
However, perceiving picture and distractor demands
simultaneous processing of both items. Therefore, naming
latencies vary as a function of the relation between the target and
the distractor (cf. Schriefers et al., 1990; Roelofs, 1992).
Concerning lexical access in language production, a twofold
picture regarding interference can be drawn: semantically, it is
known that a high number of lexical neighbours causes
interference (Schriefers et al., 1990; Levelt et al., 1991). Semantic
competitors inhibit each other, i.e. semantic closeness renders
lexical retrieval difficult, reflecting the need for deeper processing
(gathering information to select the target item). This interference
effect is supposed to be lexical, because it disappeared when
participants had to categorise the pictures instead of naming them
(Schriefers et al., 1990). The phenomenon can be assumed to
reflect competition among features of lexical items (Schriefers et
al., 1990; Roelofs, 1992; Starreveld & La Heij, 1995, 1996).
On the contrary, the presence of phonological neighbours has
been shown to facilitate processing. This phonological facilitation
effect, however, occurs because of specifying similar motoric
patterns and activity; hence, the target reaches the activation
threshold faster (Schriefers et al., 1990; Levelt et al., 1991).
“Interference paradigms study how perceptual events interfere with

each other [emphasis by the author] by virtue of similarity. The basic
idea here is that some form of interference must result when perception
and action make use of the same codes at the same time. Action should
then be impaired by concurrent perception (sic!), and perception
should be impaired by concurrent action. This prediction is not
compatible with separate coding and should therefore provide a strong
test of feasibility of the common coding approach”. (Prinz, 1997)
Although the above quote concerns action control and planning,

Prinz (1997) admits that similar views have emerged for language
perception and production [p.133]. Concerning the latter, common
coding refers to shared representations of grammatical features in
different modalities, like gender values that are the same in

production and comprehension or – of crucial importance here –
the regularity of verbs. Reliable results were obtained for gender
(Schriefers, 1993), word class (Pechmann & Zerbst, 2002) and
DC/CC (Bordag & Pechmann, 2009).
For the current study, there are manifold reasons that could
lead one to believe that the experimental paradigm was not
sensitive enough. To claim that the paradigm was not sensitive
enough means that distractors were not processed deeply enough
to activate the grammatical feature at test. For the proper
processing of distractors it is important that
(1) the target and the distractor meet at exactly the right
moment (SOA)
(2) sufficient attention is paid to the distractors
(3) linguistically, the level of the picture and the distractor
processing are the same (lemma vs. word form)
(4) distractors activate their regularity feature
Judging from personal experience conducting picture‐word

interference experiments, the SOA time windows used in the
present study were narrow enough to at least discover a trend of a
congruency effect if it really laid between two SOAs (Pechmann &
Zerbst, 2002; Pechmann et al., 2004). Additional support comes
from the fact that Bordag and Pechmann (2009) presented the
target and the distractor simultaneously, i.e. at SOA 0 exploring
inflectional processes, and got significant results for a comparable
phenomenon43.
In fact, participants processed the distractors properly as
naming latencies were affected by the type of distractor.
Comparing experimental and control conditions, naming was
fastest with neutral distractors, slower for identical distractors and
slowest when congruent or incongruent distractor verbs appeared.
43 By experience, the time window from SOA ‐200 to SOA +100 was broad enough
to ensure simultaneous processing of the picture and the distractor because
reading might be faster than production.
Thus, the clear distinction between control and critical conditions

proves the perception and processing of distractors.
In picture‐word interference tasks, irrelevant input has to be
suppressed. In discrete two stage models like the Levelt model
(Levelt, 1989; Levelt et al., 1999), only one lemma (the target
lemma) will be selected and processed on lower levels. Therefore,
it is a legitimate question whether or not distractors are processed
on the word form level. An answer is given by WEAVER++
(Roelofs, 1997): Word form perception activates lemmas and word
forms in parallel. WEAVER++ has an implemented input control
and can handle stroop‐like tasks. Hence, for the current study, it
can be assumed that distractors affect the lemma as well as the
word form level. Consequently, all lemma‐associated grammatical
features should be activated and interfere in case of incongruity.
It is not only necessary that target and distractor are
theoretically processed on the same level but the distractors’
grammatical features need to be activated properly. Although no
definite claims can be made, it is likely that distractors in a
picture‐word naming task undergo the same operations like the
target (also assumed in Bordag & Pechmann, 2009). Putatively,
distractors are processed in the same way as the target. That
means for the current experiments, that distractors are
automatically, unconsciously inflected for third person singular
like the target. In consequence of the inflectional processes, it
follows that the distractors should have activated their abstract
regularity feature. Crucially, both target and distractor are
processed on the same levels and have activated their regularity
features which were either congruent or incongruent with
different implication for the expected interference effect. So far,
there is no argument against the sensitivity of the chosen
paradigm.
7.6.2 A caveat
The encoding of the regularity feature is probably so tightly

intertwined with morphological processing that it encounters
difficulties in dissociating lemma level and word form level
processes. Morphological processing, e.g. the concatenation of

stem and affixes of regular verbs, is assumed to proceed at the
word form level but the information about a verbs’ regularity
status is part of the lemma. In the current study, it is possible that
one process overrode the other. The following paragraph will
argue that previous studies like Bordag and Pechmann (2009)
clearly showed lemma level effects with the PWI paradigm which
is not manageable with regard to German regularity.
The PWI paradigm has successfully been used to elicit
congruency effects in Czech through the activation of declensional
class of nouns and conjugational class of verbs, respectively.
Incongruent distractors delayed the articulation. Moreover,
Bordag and Pechmann (2009) set up a PWI experiment to
elucidate the process behind the DC congruency effect (see section
4.4) to explore whether the observed congruency effect originated
on the lemma or word form level. The data favour an
interpretation of the congruency effect occurring not on the form
level. Instead, the congruency arose on the lemma level because
the same influence of declensional class was observed regardless
of whether target and distractor were overlapping in form or not.
Therefore, it is concluded that abstract features competed for
selection. DC is a grammatical property stored in the lexical entry
which has morphological implications like regularity. The
properties of Czech conjugation, however, did not allow
constructing an equivalent experiment with verbs.
For German regularity, an experimental dissociation of lemma
and word form processes remains unresolved. From the
experimental design employed in Experiments 1‐4, it is not clear
whether lemma and word form level processes interfered with
each other and if they probably cancelled each other out. The
German conjugational system does not allow a clear dissection of
morphological processing and encoding of grammatical features,
nor does the declensional system. As a last resort one could try to
elicit regularity congruency effects with a non‐verbal variant of
the PWI paradigm. As the task is non‐verbal, no word forms were
involved and the effects could be attributed to the lemma level.
On a more practical note, experimental possibilities in speech

production are limited by methodological constraints. Artefacts
produced by vocalizations present a great challenge. In particular,
the PWI paradigm taps comprehension and production processes
in contrary ways. Further research is needed to investigate the
influence of comprehension and production on behaviour in PWI
experiments.
7.6.3 Intermediate conclusion
The purpose of the experiments reported above was the

investigation of the representation of a regularity feature and its
time course of activation.
Unfortunately, the critical manipulation failed, since one
cannot conclude anything from null effects. However, it is
reasonable to further follow up on the assumption that regularity
is a lexically specified grammatical feature. Two reasons are:
Regularity is an arbitrary feature that has to be learnt and
stored (like gender)
According to the WR both regular and irregular forms are
stored in the same mental lexicon and have to be retrieved. The
language production system needs information on when the
regular route has to be blocked and when it has to be open.
Morphological processes have to be triggered.
So far, the data allow only speculations on how regularity as a
property of verbs is represented: as the regularity status did not
induce morphological interference, it could be represented as a
privative feature. Privative features are known from phonology
like for example [labial] to describe the place of articulation of
phonemes. Privative features are either existent (e.g. [labial]) or
absent. It becomes evident that privative features do not have
counterparts (e.g. [‐labial] and [+labial]) which could cause
competition. Regarding regularity, let us assume that only regular
verbs are provided with a feature specifying it for regular. As there
is only one value for the feature, no competition could arise in the
experiments and therefore no congruency effect was observable. It
could be that there is no node representing the verbs regularity
status but that regular word forms are marked by a slot for the
tense suffix which irregulars are lacking. However, this answer is
highly speculative.
There is no denying of the robust regularity effect observed in
the experiments reported. Regular verbs are produced
significantly faster in past and present tense than non‐regular
verbs. Neither previous studies nor Experiments 1‐4 considered
whether regularity is represented as a property of individual
forms (implicit assumption of DRM) or of whole inflectional
paradigms. Hybrid verbs are suitable to test these two hypotheses.
This issue is addressed in Experiment 5 once more. To explore the
relation between verb type and tense and to figure out processing
mechanisms, the experimental paradigm can be simplified.
7.7 Experiment 5
The previous experiments tested whether an interference effect for

regularity can be discovered and located on the lemma level. They
served to approach an answer about storage and retrieval of verb
lemmas. Even though there is a feature, it is still an open question
why non‐regular verbs are produced more slowly than regular verbs
(see section 2.1 and Seidenberg & Bruck, 1990 as cited in Seidenberg,
1992) and whether regularity is represented as a property of
individual forms or of whole inflectional paradigms. Therefore, the
purpose of Experiment 5 is to address morphological processing
issues. Attention is shifted from the lemma level to the word form
level and more precisely to the morphological layer.
The widespread view to explain the distinction in articulation
latencies of regular and non‐regular verbs is the blocking mechanism
(see section 2.3.1). The Dual Route account is now challenged by the
threefold distinction of verbs in German in regular, hybrid and
irregular classes. Hybrid verbs are the object of investigation in the
following experiment. It was a simple picture naming paradigm
studying inflectional processes from articulation latencies in present
and past tense. If articulation latencies reflect different mechanisms
the Tense by Regularity interaction and a comparison of processing

hybrid verbs to regular verbs are of crucial interest.
For hybrid verbs, the Dual Route Model predicts a qualitative
distinction between a rule‐governed process in present tense and
lexical retrieval of ready‐made entries in past tense and no
distinction between the present forms of hybrid verbs and regular
verbs.
Alternatively, in a lexical entry complexity account, comparable
to Lukatela (1980) for language comprehension, it is reasonable to
expect the same reaction times for irregular and hybrid verbs in both
the past and present tense because both have to select the correct
word form out of several possible word forms. Since regular verbs
do not have competing word forms, they are produced faster than
the former. Furthermore, the expectations include a Tense effect for
irregular and hybrid verbs (past slower than present) but not for
regular verbs. The reason is that for regular verbs, the same stem is
employed in present and past tense.
In fact, Experiment 5 is a follow up of Experiments 1 to 4.
Articulation latencies measured in Experiment 4 showed no Tense
by Regularity interaction, but instead a main effect of hybrid and
irregular verbs opposed to regular verbs. This finding is
contradictory to the assumptions of the DRM. For elucidation and
replication, a simple picture naming experiment is sufficient. If it
succeeds to replicate the previous findings and to find an interaction
between the factors Regularity and Tense, the blocking mechanism
has to be reconsidered.
7.7.1 Method
Participants
Thirty‐six native German speaking students of the University of

Leipzig were recruited. They were paid as compensation for their
participation.
The same set of 27 black and white line drawings from Experiment 4
was used as experimental stimuli. Pictures depicted the actions of
intransitive German verbs. Each of the verbs was classified as
regular, hybrid or irregular verbs. The groups were matched for
frequency and onset.
Participants were first administered two familiarisation phases. In
the first level they saw the stimuli with the correct name of the
picture written below in the infinitive. For practice they were asked
to produce the infinitive for each target picture. The infinitive was
chosen to minimise word form repetitions and to avoid pre‐
processing before the experimental measurement itself. A second
familiarisation level checked whether participants memorised all
picture names correctly from memory without picture names written
on the screen.
At test, each trial started with a fixation‐star appearing in the
centre of the screen that indicated that the initial sentence fragment
„jemand“ [someone] had to be produced. After the disappearance of
the fixation star the picture emerged in the middle of the screen. The
depicted action was named in the 3rd person singular present or past
tense. Articulation latencies were measured by a voice key. Tense
was counterbalanced across subjects and blocked. Participants were
instructed which tense was required before each of the two blocks.
Each picture was named only once per tense to avoid effects of
repetition priming.
Apparatus
The setting was the same as used and described for the picture‐word
interference experiments.
7.7.2 Results
Incorrect responses were excluded from all analyses. Reasons for

exclusion were wrong namings, hesitations during articulation,
technical problems with measurements and voice key triggering
(exclusion of 13.3 % of the data for these reasons). A cutoff discarded

articulation latencies shorter than 200 ms and longer than 1500 ms as
well as values plus or minus two standard deviations from the
subject’s mean. They were handled as missing values. The total of
missing values was 13.5 %.

Cut off 0.2 0.1 0.2
Reaction times measured in Experiment 5 can be seen in Table 23.

deviations in parentheses), Varied by Regularity and Tense (Experiment 5).
Regularity
Tense irr hyb reg M
638 619 480 577
past
(185) (171) (106) (171)
577 547 447 520
present
(170) (158) (93) (153)
606 583 462 549
M
(180) (169) (101) (165)
A two‐way repeated measures analyses of variance (ANOVA)

yielded significant main effects of REGULARITY (three levels),
[F1(2,70) = 116.61, MSE = 3579.28, p < .001; F2(2,16) = 21.61, MSE =
5312.15, p < .001] and TENSE (two levels) [F1(1,35) = 41.41, MSE =
4058.29, p < .001; F2(1,8) = 168.14, MSE = 254.98, p < .001]. The
interaction between REGULARITY and TENSE reached significance
by subjects and very scantily by items [F1(2,70) = 6.92, MSE = 1153.38,
p < .01; F2(2,16) = 3.64, MSE = 485.99, p = .05]. A post hoc Scheffé‐Test
(diffcrit; p < .05 = 20.2) revealed that reaction times do not differ between
irregular and hybrid verbs and that their articulation latencies
depended on the factor tense (production is faster in present tense)
whereas that particular tense effect is not significant for regular
verbs.
Analyzing error rates as independent variable for the interaction
between the two factors Regularity and Tense all missing values,
including wrong namings, hesitations, time‐outs and cut offs, were
taken into consideration (13.5 % Table 21 above).
Table 24. Percentage of missing values, varied by Tense and Regularity
(Experiment 5).
Present Past total

regular 3 8 5,5
hybrid 17 13 15
irregular 19 21 20
total 13 14 13,5
The analyses of error rates mirrored those of reaction times with

respect to the main effect for Regularity [F1(2,70) = 33,88, MSE = 0.01,
p < .001; F2(2,16) =10.06, MSE = 0.01, p < .01] and the interaction
between Tense and Regularity for subjects only [F1(2,70) = 5.11, MSE
= 0.01, p < .01; F2(2,16) = 2.04, MSE = 0.01, p = .162]. A post hoc
Scheffé test (diffcrit; p<.05. = 0.09) resolving the Regularity main effect
revealed that non‐regular verbs were associated with more errors.
7.7.3 Discussion
Experiment 5 measured articulation latencies for picture naming of

actions. Three groups of German verbs – regular, hybrid and
irregular – were explored in the past and present tense. Postulating
two different mechanisms for the processing of regular and irregular
inflection, the DRM cannot account for all data in Experiment 5, in
particular not for the fact that even regular forms of hybrid verbs
(present tense) are produced more slowly than regular forms of the
regular verbs (past and present tense). As the data did not show an
interaction with tense for hybrid verbs, it is necessary to revisit the
DRM.
Crucially, the present forms of both hybrid and irregular verbs
were produced more slowly than (and did not differ statistically
from) the regular present forms of the regular verbs. Irregular and
hybrid verbs were produced more slowly in the past than in the
present tense. Present tense and past tense reaction times did not
differ for regular verbs. Since the material has been checked carefully
for word form and lemma frequency as well as for other
confounding factors, the interaction can only be attributed to a
deeply rooted linguistic phenomenon.
The naming latencies of hybrid and irregular verbs did not differ
from each other in both tenses. Strikingly, verbs according to their
articulation latencies can be split up and subdivided into regular
verbs and verbs having at least one irregular form (hybrid verbs and
irregular verbs). This consideration goes hand in hand with the two
groups as revealed by the Scheffé test. The crucial explanatory factor
seems to be the complexity of the lexical entry: if a verb has multiple
stems (irregular and hybrid verbs), the retrieval of the appropriate
one takes longer than the retrieval of a single stem entry (regular
verbs). Regular verbs do not have competing word forms. If for
irregular and hybrid verbs two or more word forms are encoded,
then one may be selected during morphological encoding. The
selection of the current word form is time consuming. Error rates
confirm the higher processing cost for irregular and hybrid verbs
since more errors occurred for naming of hybrid and irregular verbs
independently of tense.
The regularity effect observed in Experiment 5, showing a
dissociation between regular and non‐regular verb forms, is
supposed to occur even without sentential context, as it is caused by
the lexical entry complexity that should not change by syntactic
context. In general, the results of Experiment 5 are in line with the
previous experiments (1‐4): The Regularity effect was observed with
both the PWI paradigm and the naming task. Comparing the data of
the picture‐word interference task and the simple naming, the Tense
by Regularity interaction might not have been significant in

Experiment 4 because the Tense effect was not as pronounced as in
Experiment 5. It is possible that the tense effect was evened out in
Experiment 4 because of too many item repetitions.
Hybrid verbs’ present tense articulation latencies show that
regularity cannot be bound to individual word forms as they are
superficially regular but dissociated from regular verbs in reaction
times. Hence, regularity generalizes on all forms within a paradigm.
From a learner’s point of view, it is really pragmatic to store and
mark verbs paradigmatically either as regular or non‐regular. If there
is at least one non‐regular form in the paradigm, verbs are learnt and
treated as non‐regular. That makes sense from a learner’s
perspective.
The following chapter concludes this thesis along these lines and
discusses an adaptation of WEAVER++ (Roelofs, 1997). So far, the
model is neutral about the question of whether one or two word
forms are encoded for non‐regular verbs. Possibilities for further
research are discussed in the General Discussion (section 8).
8 General Discussion
This dissertation investigated the representation of regularity and
the inflectional processes involved in the production of regular and
non‐regular German verbs. Particularly, the generation of past and
present tense stems was explored. So far, the role of regularity has
only been explored with respect to past tense or past participle
forms, where it is overtly marked in most investigated languages
(Pinker, 1991, 1999; Clahsen, 1999; Clahsen et al., 2001). The data
presented dispute the Words and Rules Theory (Pinker, 1999) and
the underspecified lexical entries assumption by Clahsen (1999). The
data are consistent with a model that takes lexical entry complexity
into account where competing word forms of a verb determine its
processing difficulty.
The present thesis explored three groups of German verbs in the
past and present tense: regular, irregular and hybrid verbs. Hybrid
verbs in German have completely regular conjugation in the present
tense and irregular conjugation in the past tense. The first four
experiments aimed at tapping into the activation and encoding of
abstractly represented regularity information in speech production.
The experiments used the traditional picture‐word‐interference
paradigm to test for a regularity congruency effect. Based on the
conjugation class congruency effect in Czech (Bordag & Pechmann,
2009), it was hypothesised that regular and non‐regular verbs should
interfere more strongly across category than two verbs of the same
verb type. Slower RTs in the incongruent condition than in the
congruent condition were expected, reflecting competition between
abstract grammatical features for regularity. Contrary to this
assumption, the critical conditions did not exhibit the expected
congruency effect. Nevertheless, all four main effects in Experiments
3 and 4 (for distractor type, regularity, tense and SOA) were
141
142 CHAPTER 8: GENERAL DISCUSSION
significant. Most importantly, the data showed an effect of distractor

type. Although this effect was obtained only between experimental
and control ‐ not within experimental – conditions, it allowed the
conclusion that the paradigm employed was sensitive to the
experimental manipulation. The null results in Experiments 1‐4
tentatively speak against an abstract representation of regularity as
generic nodes.
Neither previous studies nor Experiments 1‐4 considered the
possibility that regularity may be a property of individual forms
(implicit assumption of DRM) or of whole inflectional paradigms.
Hybrid verbs are suitable to test these two hypotheses because their
regularity status dissociates in past and present tense. Thus, a picture
naming task was used in Experiment 5. This task revealed that
regular verbs were produced significantly faster than hybrid and
irregular verbs. The significant interaction of regularity and tense
showed that the past tense naming latencies of non‐regular verbs
exceeded their present tense naming latencies. However, for regular
verbs, past and present tense naming latencies were similar.
Considering the absence of the regularity congruency effect in
Experiments 1‐4, it seems to suggest that regularity is not – or at least
not alone – a generic feature like e.g. conjugation class, although very
similar experiments on conjugation class in Czech (Bordag &
Pechmann, 2009) supported the idea of a generic feature. A clear
congruency effect – longer articulation latencies with differing
conjugational classes – in a PWI task was in accordance with generic
nodes for verb classes. Bordag and Pechmann (2009) therefore
suggested the existence of psychologically real representations of
conjugational classes in Czech. The corresponding node for
conjugational class is selected for each verb appropriately during
grammatical encoding. Abstract class nodes mediate the selection of
the appropriate ending on the phonological level. In addition,
Bordag and Pechmann ensured that the congruency effect does not
reflect competition for the phonological form of the appropriate
ending. In discrepancy to Bordag & Pechmann (2009), the results
presented here did not differ for items paired with distractors
sharing the regularity status than for items paired with distractors
deviating in regularity.
143
Comparing naming latencies of both studies at the same level of

SOA (0 ms), participants were naming action pictures in Czech at an
average of about 100 ms slower in all conditions than German
participants. This could be an effect of different subject groups.
Alternatively, it may also be that the phonological properties of
Czech words require more processing time. Czech permits more
complex consonant clusters and has alveolar trills which do not exist
in German words. Most syllables in Czech start with consonants (up
to four). Czech has also relatively infrequently open syllables of the
structure (CV). Preparation and realisation of phonetic plans of
consonant clusters take time and cause longer latencies which in turn
set up time and ample opportunity for interference to occur.
However, the present study used the same method; the results and
the evaluation of the paradigm (section 7.6) argue against an isolated
methodological cause.
Theoretically, there are parallels between CC in Czech and verb
regularity in German. Both are features determining the inflection of
verbs. However, the data presented above argue against such an
assumption. German verbs seem not to be organised in classes or,
more likely, the classes are not mentally represented in the form of
generic nodes as was predicted. Hence, the German conjugation
system deviates from the Czech conjugation system on some major
points. Although German verbs’ regularity status classifies and
clusters verbs into some kind of classes, these classes are different
from Czech CCs. This raises the question of which properties of
Czech verbs may account for this difference? Most probably, the
experiments explored different levels of representation. CCs in
Czech are a subordinate ordering principle and exceptions (i.e. non‐
regular forms) can be found within CCs. In German the reverse is
true: German verbs are primarily regular or non‐regular and non‐
regulars form smaller sub‐classes (Wiese, in prep.).
Referring to Bordag and Pechmann (2009), regularity is an
internal feature like gender or CC in Czech. All three are lemma‐
specific, context independent, have morphological implications and
are indispensable features. Regularity cannot be bypassed or
neglected in production as could be done for gender (e.g. La Heij et
al., 1998 reporting absent gender congruency effects on Dutch bare
noun naming). Like CCs, regularity is not dispensable, because in

order to inflect a verb it is first necessary to know which processes to
start. These attributes have implications on whether and how
regularity is represented. The lack of evidence for explicitly
represented regularity information in the mental lexicon constitutes
a puzzle: which mechanism triggers the decision to handle a stem as
either regular or non‐regular? Since regularity is an arbitrary
feature44, it can only be lexically organised. As the regularity status
did not induce morphological interference in the experiments
reported above, it seems to be inherently represented with the
structure of the lexical entry of each verb.
For combining these conclusions with the research cited in
chapter 1.3 about typologically different inflectional systems, it is
helpful to reconsider the typological dimension. One frequently
applied criticism challenging generativist explanations for linguistic
phenomena is that it may not be possible to find universal
mechanisms of lexical organisations across languages (cf. Lukatela et
al., 1980). Lukatela et al. (1980) investigated the lexical representation
of Serbo‐Croatian nouns in a lexical decision task and found
considerable differences between Serbo‐Croatian and English. Both
are Indo‐European languages and therefore related to German.
English and German belong to the Germanic language family and
Serbo‐Croatian is a member of the South‐Slavic branch. As Czech
also belongs to the Slavic branch, German and Czech participants do
not need to behave in the same way in similar tasks.
Even though the experimental paradigm did not reveal the
expected congruency effect, it yielded reliable regularity main effects
in Experiments 2‐5. The analyses revealed that regular verbs in the
present and past tense were produced significantly faster than all
other verbs. Crucially, the naming latencies of hybrid and irregular
verbs did not differ in both tenses. Indeed, the results reported here
44 While most German verbs are regular, many of the most frequent verbs are non‐
regular. Contrariwise, almost all low‐frequent German verbs are regular and all new
verbs are conjugated regularly. Though it is a really interesting question whether
frequency or pragmatic factors impose pressure on the vocabulary or the language
processor, this work cannot contribute to this issue (the reader is referred to Schmidt,
Langner Helmut & Wolf, 1996: 191–203, 241–253, 309–323).
145
for irregular verbs converge with those obtained in previous

experiments. Prasada et al. (1990) and Seidenberg & Bruck (1990)
tested whether regularity affects articulation latencies and suggested
Dual Route processing and connectionist networks, respectively. As
mentioned in the introduction, this thesis is not foremost a
contribution to the past tense debate (e.g. Prasada et al., 1990;
Seidenberg & Bruck, 1990). Nevertheless, the reaction times to
German hybrid verbs in the present tense are a challenge for single
route models. As in the present tense, hybrid verbs are supposed to
share features with regular verbs, both should be produced equally
fast. In fact, since hybrid verbs are produced as slow as irregular
verbs, the obtained results are in conflict with single route models as
well. Similarly, the past tense results reported converge with those
obtained in previous experiments designed to test Dual Route
Models (Clahsen, 1999; Clahsen et al., 2004; Marcus et al., 1992) in
that they demonstrate the dissociation of regular and non‐regular
verbs. Concerning the Words and Rules Theory, a crucial flaw is that
many aspects cannot be easily applied to the processing of German
and especially German present tense verbs.
According to the Words and Rules Theory (Pinker, 1997, Pinker,
1999; Clahsen, 1999), regular forms are composed of a stem and
suffixes whereas irregular forms are stored as ready‐made entries.
Regularity is represented as a property of individual forms (implicit
assumption of DRM). To produce non‐regular verbs, the regular
route is pre‐empted by the non‐regular route. This is called blocking
and delays the production of non‐regular verbs. The important
conclusion to be drawn from criticisms against the blocking
mechanism (section 2.3.1) is that German provides a test case for the
blocking mechanism if one is to consider not only past tense, but also
present tense. In German present tense, (er) beiß-t [he bite‐s] and (er)
lach-t [he laugh‐s] are both built regularly. However, beißen [to bite]
is non‐regular and lachen [to laugh] is regular in the past tense. The
crucial finding of longer reaction times for hybrid verbs than for
regular verbs in the present tense in the current study is challenging
for both Pinker (1999) and Clahsen (1999): Both claim fully regular
processing for regular words and do not differentiate between
present and past tense. However, it was found in the present thesis
that the superficially regular inflection of hybrid verbs affects

production latencies equally strong as the irregularity of verbs. In
short, neither blocking nor Dual Route accounts can fully explain the
data presented here.
Clahsen (1999) did not consider present tense processing and
therefore no claims about present tense are made in his work. It will
be shown in the following that non‐regular verbs’ lexical entries are
not only underspecified for past tense word forms but also for
present tense. Janssen (1999) found initial evidence for irregularity in
present tense in Dutch; the current study underlines the validity of
his claim.
Still, the following perspective provides a more promising line for
the interpretation of the current results: irregular and hybrid verbs
have more than one potential stem. According to this criterion, verbs
can be split up and subdivided into regular verbs on the one hand
and non-regular verbs having at least one irregular form (hybrid
verbs and irregular verbs) on the other hand. The classification of
German verbs in two instead of three classes of verbs goes hand in
hand with the current psycholinguistic results. Even though a rule
could be used to generate present tense forms, hybrid verbs follow
processing principles of non‐regular verbs (assuming articulation
latencies reflect processing mechanisms; Donders, 1868, 1969).
Hence, no individual forms of hybrid verbs are classified as regular
or non‐regular, but rather the whole paradigm of a particular verb.
As this distinction has not been made before, this had to be shown
first. The crux of the matter: their typical and common characteristics
are stem vowels like /ei/ in beißen and /ie/ in biegen that do not have
fronted counterparts and therefore cannot mutate to umlaut. Their
nature is irregularity and so they are listed together with irregular
verbs in German dictionaries (Kunkel‐Razum, 2006).
Non‐regular verbs are intrinsically marked as non‐regular. They
are psychologically distinct from regular verbs in that their lexical
entries are more complex. More precisely, regular and non‐regular
verbs’ lexical entries differ concerning their complexity at the level of
the word form. Processing proceeds in similar ways at the lemma
level, it is likely that the origin of the latency differences can be
located on the word form level where for non‐regular verbs multiple
147
stems are available. These stems are not yet inflected for person and
number.
Consequently, the aspects emphasized allow the conclusion that
if a verb has multiple stems (irregular and hybrid verbs), the
retrieval of the appropriate one takes longer than the retrieval of a
single stem entry (regular verbs). Although hybrid verbs are
superficially regular in the present tense, the present experiments
demonstrated that they form a single group with irregular verbs. It
can be seen in Figure 7 that non‐regular verbs have several word
forms available to choose from.
L
E
M lachen beißen geben ?
M
A
W Selection
O lach- beiß- biss- geb- gib- gab-
R
T
+t +t +t Inflection
F
O
R
M
Figure 7. Complexity of lexical entry. Simple and complex German verbs.
It is conceivable that it is not blocking that prolongs the reaction times

for non‐regular verbs, but an additional stem selection process. The
following assumptions cannot be directly drawn from the data, but
are in the scope of interpretation. The regularity effect is indicative of
a selection effort in the non‐regular inflection. The generation of the
correct word form is more costly for non‐regular verbs because more
stems are related to one lemma. Ranging from two to five potential
word forms (including subjunctive), non‐regular verbs can posses a
veritable pool of stems, e.g. brechen [to break]: brech‐e, brich‐st, brach‐
∅, ge‐broch‐en, bräch‐e. Hence, compared to regular verbs, there is
selection in the access of non‐regular verbs as opposed to the mere
lexical retrieval of a single stem from a single lemma. Most
importantly, this argument against blocking is neither a descriptive‐
linguistic nor a deductive, but a cognitive one. Having more than
one potential stem in the lexical entry, processing costs for non‐
regular verbs are higher because of the selection of the appropriate
stem. The selection mechanism dispenses with the blocking
mechanism and gets around the impediment of the general principle
of mental grammar to creatively compute complex utterances ‐ an
extremely powerful device for communicating.
The lexical entry complexity model proposed here has substantial
similarities with the Satellite model of Lukatela (Lukatela et al.,
1980), although the Satellite model was initially conceived for the
representation of nouns. For these, the model locates the nominative
as the base form in the centre, inflected word forms (satellites) are
arranged as nuclei around it. A parallel architecture can be
envisaged for the representation of German verbs. Potential verb
stems (e.g. brech-, brich-, brach-, -broch-, bräch-) are ordered around a
base form, e.g. the infinitive (e.g. brechen [to break]) or the first
person present tense (e.g. (ich) brech-e [(I) break])45. As regular verbs
only have one stem, they are atomic. Crucially, the more satellites a
verb has, the more difficult it is to retrieve its correct word form.
However, the Satellite model is a very uneconomic variant of lexical
representation. Verb paradigms in German have more individual
forms than nouns do (six per tense and mood vs. four cases per
singular and plural), so a full representation of inflectional
paradigms produces a lot of redundant information46. Furthermore,
as a consequence of full listing, one would expect numerous
suppletive forms in any language lacking German verbal inflection
apart from sein [to be] and haben [to have]. The model proposed here
will account for economy principles and subregularities within non‐
regular inflection.
The representation of non‐regular verbs is best explained and
modelled in terms of underspecification of lexical entries. In
theoretical linguistics, underspecification is a phenomenon where
certain features are not represented in the underlying structure, but
45 Lukatela’s argument for exposing the nominative in the centre were significantly
shorter reaction times to that particular form. I am not aware whether the infinitive of
any other inflected form is faster than others. To elaborate the Satellite model for
verbs, all inflected forms would have to be tested to find the base form.
46 Except for suppletive paradigms like sein [to be].
149
are unequivocally derived in an incremental process. This proposal

is consistent with Clahsen (1999), but will be extended to present
tense. Underspecification theories (e.g. Minimalist Morphology, cf.
Wunderlich, 1996; Clahsen, 1999; Distributed Morphology, cf. Halle
& Marantz, 1993) hold that features should only be underspecified if
their values are predictable. For example, English front vowels (/i, ɪ,
e, ɛ, æ/) are unrounded; hence, it is not necessary for these phonemes
to include the distinctive feature [−round], because all [−back]
vowels are [−round] vowels. This principle inspired the basic
architecture of lexical entries for non‐regular verbs in Clahsen’s
(1999) approach: it assumes internally structured lexical entries in
form of feature pairs of phonological and morphological
information. However, the new conception of structured
underspecified lexical entries is characterised by omission of features
in the mother node (cf. Clahsen, 1999, underspecified subnodes).
Subnodes are incrementally specified instances of the mother node.
Mother node and subnodes are complementary templates. The
mother node provides one or more slots which have to be filled by
one of the related subnodes. Slots are filled while generating the
intended word form. Information is not inherited but mother node
and subnodes complement another. Compare the lexical entry for
the irregular verb singen [to sing], as shown below:
s_ng-
[ …i …]pres [ …a …]past [ …u …]part

Figure 8. Lexical entry of the hybrid German verb singen [to sing].
The tense‐feature is conceptually specified. A subnode is selected

according to the specific tense value. With the selection of the
appropriate subnode, the missing vowel can be inserted in the
mother node. Non‐regular verbs mainly exhibit vowel changes.
Selecting the subnode takes time and explains the regularity reaction
time effect as this cost does not apply to regular verbs. Subnodes are
instantiations of the mothernode, which are inserted at the level of

word form.
Nevertheless, the distinction between regular and non‐regular
verbs still requires the corresponding distinction of two qualitatively
different morphological processes (cf. Pinker, 1999, and others). The
data presented here showed that regular and non‐regular verbs were
differently affected by the tense of an utterance. The interaction
effect observed in Experiment 5 demonstrated a slowing down in
past tense naming of irregular and hybrid verbs but not of regular
verbs. Regular verbs take only one stem in both tenses and add the
appropriate suffixes. As regular verbs employ the same stem in
either tense, no RT difference is expected. Non‐regular verbs,
however, employ special past tense stems. These are used
infrequently because of the generally low discourse frequency of
past tense in German (see chapter 5: 15 percent past tense
utterances). Highly frequent subnodes (e.g. present tense) are
activated much easier than low frequent ones (e.g. past tense). The
observed tense effect traces back to this fact. Thus, the tense effect of
non‐regular verbs may be a hidden frequency effect and also
confirms two different morphological processes for regular and non‐
regular verbs. Still, in a modified version, the Dual Route Mechanism
can capture German verb production. To produce a regular verb, the
tense suffix is activated together with the stem, i.e. the null
morpheme in present tense and –te for past tense. Stem and suffix
need to be concatenated subsequently. Hence, regular word forms
are computed using two morphemes (word form and tense suffix)
and the operation merge. This mechanism resembles the one
proposed by Pinker (1999).
The assumption of separate lexical entries for each non‐regular
form has to give way to the idea of underspecified lexical entries of
non‐regular verbs (see Discussion of Experiment 5; Clahsen, 1999).
Lexical entries of non‐regular words do not contain full word forms
because this would multiply shared information (e.g. s, n, g, in the
example above), which would be quite redundant. German features
much regularity among non‐regular inflection. Some researchers
even assume subclasses like sinken-sank-gesunken and trinken-trank-
getrunken (Wiese, in press). The subnodes of underspecified lexical
151
entries introduced above may visualise the relatedness of subclasses

of verbs: sinken-sank-gesunken and trinken-trank-getrunken share the
same subnodes. Subnodes are finally inserted in different mother
nodes. The structure clearly prohibits full listing.
It is important that previous psycholinguistic data can be
explained with the lexical entry complexity account. Relevant
questions here are: (1) How, for example, is it possible to explain the
anti‐frequency effect of regular verbs (Clahsen, 1996) without
blocking?; (2) Why do children over‐generalise the formation of
regular verbs by applying the rule to non‐regular verb stems?; (3)
Why do most speakers of German not produce beißte as preterite of
beißen [to bite]?; (4) Why are new words in a language inflected
regularly?
The answer to the first question is that high frequent regular
verbs’ lexical entries are also complex. They embody both one
decomposed and one full form for the purpose of immediate
retrieval of very frequent forms. The appropriate form has to be
selected from several forms which takes time. The phenomenon of
over‐generalisation of children traditionally has been explained as a
phase in language acquisition of becoming aware of rules and
mastering their application. In the selection account, rules exist
further on. They have to be learnt and applied to regular verbs. At
the age in which children overgeneralise the rule to non‐regular
verbs, they have not yet learned the particular non‐regular verb
form; however, they know the concept of the verbs and its infinitive.
As there is no stored word form to retrieve, they apply the rule to the
infinitive which is available.
The past tense stem of beißen is biss. The lexical entry of beißen
contains a mother node [b_] and a subnode [… iss …]past, which is
unique in its past tense feature and can be inserted in the mother
node template. Processing is triggered by the structure of the lexical
entry. Mother node and subnode map onto each other. There is still a
Dual Route architecture but the new model stipulates rules on both
routes. New words in a language are inflected regularly by add –
suffix rules because new words do not have a lexical entry from
which non‐regular forms can be generated. Then, for example, the
past tense of a new verb is computed because the rule add –te can
apply to the new root.
Before closing this discussion, the implications of the current
results for language production models will be considered. The
theoretical considerations together with the empirical data allow for
the formulation of a comprehensive hypothesis concerning the
production of regular and non‐regular verbal inflection. In
particular, the focus is on the Levelt model (Levelt, 1989; Levelt et al.,
1999), because the hypotheses were mainly derived from its
architecture. The emphasis is on the preparation of past and present
tense regular and non‐regular verb stems for articulation. Both types
of verbs are supposed to inflect tense marked stems for person and
number alike though this is in need of further investigation. The
crucial adaptation is to integrate the selection mechanism and to
dispense with the blocking mechanism (Pinker, 1999).
Foremost, the adaptation of the Levelt model has to account for
the fact that structural complexity of lexical entries renders the
selection of correct word forms more difficult. In addition, the
adaptation has to locate the application of rules within the model.
All in all, this proposal is also a modification of the Words and Rules
Theory (Pinker, 1999) and its implementation inside the complex
production process as represented in the Levelt Model under
consideration of underspecified lexical entries.
Figure 9 shows the full model of lexical entry complexity based
on the Levelt model, but adapted to deal with both the previous and
the current data.
153
Figure 9. Generating the past tense of lachen (reg) [to laugh] and singen
(nreg) [to sing].
For present tense generation, the past feature is deactivated but the present
tense node is activated instead. Further, ‐∅ instead of –te and […i…]pres
become available. They do not start by default and in parallel, unlike
established in the Words and Rules Theory. No blocking is assumed, but
time‐consuming selection and control processes.
Tense is conceptual information and is activated during the

conceptualisation of the message. After conceptualizing the message,
corresponding verb lemmas are activated. The tense information is
represented as a diacritic feature at the lemma level (Levelt et al.,
1999). If a past tense sentence is intended, the PAST diacritic will be
activated together with, for example, the verb SINGEN. Lemmas and
diacritic values are linked to corresponding morphemes. At the word
form level, two nodes are involved in the production of regular
verbs: the root and the tense suffix. Both are concatenated by a rule
merging the constituents. Non‐regular verbs activate their complex
lexical entry. The Levelt model can handle selection at the word‐
form level. Its implementation in WEAVER++ is neutral about the
question of whether one or two word forms are encoded for non‐
regular verbs. If two forms are encoded, one of these may be selected
during phonological encoding. Non‐regular verbs activate the
template and insert the vowel specified for the intended tense: a
select & insert rule applies and rule‐driven concatenation takes place
at the word form level.
In the next step of the process, according to the revised model,
person and number suffixation brings together the processing
mechanisms of regular and non‐regular verbs, the hypothesis being
that inflection for person and number proceeds similarly for both
verb types. Suffixes for person and number are attached to the stem
in any case, according to the feature specified conceptually. A
specific rule applies to add the inflectional ending for person and
number to the tense marked stem. Phonological encoding follows.
The discussion demonstrates the interplay of representation vs.
processing. Regularity must be mentally represented information that
triggers rules. In fact, both regular and non‐regular verbs are
generated by rules. Rules are lexical procedural knowledge (Levelt,
1989:185) combining roots and morphemes to tense marked stems
and, finally, fully inflected word forms.
It is not competition, but complexity that renders correct selection
of inflected word forms more difficult. The lexical entry complexity
account predicts long RTs for inflected forms throughout non‐
regular paradigms not only for superficially visible non‐regular
word forms. Hence, it predicts the processing advantage of regular
155
verbs also for subjunctive (“Tom sagt, jemand lache.” [Tom says
somebody laughes]) and the German present tense participle (lachend
[laughing]). Subjunctive and present tense participle are conjugated
completely regularly. However, non‐regular verbs’ lexical entries are
more complex. Although all forms of subjunctive and present tense
participle are superficially regular, according to the theory of lexical
entry complexity, non‐regular forms should yield slower responses.
Moreover, the selection may be more difficult with high similarity.
Further exploration of the selection mechanism is conceivable with
suppletives in comparison to irregular verbs, the former being less
similar to each other than irregulars. Therefore, reaction times
should be shorter. Whether the inflection for person and number
proceeds similarly across regular and non‐regular paradigms is in
need of further inspection.
The research reported here may also establish important new
assumptions concerning noun phrase production. The linguistic
aspects and the proposed cognitive mechanisms might be similar for
the type of memory representation which I proposed for regular and
non‐regular verbs. Nouns in plural either have null morphemes
[Segel ‐ Segel], umlaut [Vater ‐ Väter], Suffix [Auto ‐ Autos] or Suffix
plus umlaut [Baum ‐ Bäume]. Hence they differ in their number of
possible stems. A successful replication of the observed effects with
nouns may account for a general cognitive mechanism and could
tease apart the influence of person and number encoding. If the
effect shows up in singular where no umlauting occurs, it can be
concluded that it is not a phonological process (cf. Penke, 2006) that
delays naming. Then, the processing disadvantage of non‐regular
nouns can only be attributed to the selection of one stem out of
multiple stems.
One consequence of a class‐wise representation of verbs with
respect to agrammatic aphasic patients is that if they have a disorder
in the generation of verb forms, either because of an inflectional or a
lexical retrieval deficit, the production of inflected verbs should be
impaired in either tense, not only in a particular one.
The experiments presented in this thesis investigated the

representation of regularity and the inflectional processes involved
in the production of regular and non‐regular German verbs. Three

groups of German verbs were explored in past and present tense to
address these questions. Existing models did not explicitly address
the question how regularity is represented and whether regularity is
represented as a property of individual forms (implicit assumption
of DRM) or of whole inflectional paradigms. The experiments
investigated hybrid German verbs to test these two hypotheses. The
results of the current experiments do have important consequences
for the Word and Rules Theory mentioned in the theoretical part of
this thesis (Pinker, 1999). In sum, the present thesis provides data
supporting that regularity is paradigmatically, inherently coded in
lexical entries. Rather than through abstract node representation, the
paradigmatic effects can be explained as a resulting from complexity
of lexical entries. For defining complexity, it has to be differentiated
whether a verb has one or multiple stems. If a verb has multiple
stems (irregular and hybrid verbs), the retrieval of the appropriate
one takes longer than the retrieval of a single stem entry (regular
verbs). In other words: the complexity account outruns the blocking
mechanism of the Words and Rules Theory (Pinker, 1999).
Importantly, shedding light on the past and present tense of German
verbal inflection, this thesis gains a more complete picture of mental
lexicon and grammar. Still, at this point, it is only possible to
complement the speech production model of Levelt (Levelt, 1989,
1999) along the temporal axis. It will be promising to see whether
there exists a mechanism of language control for the selection of one
particular stem or the suppression of remaining stems, respectively.
Last but not least, future research will have to investigate
neuroanatomical correlates of selection processes and whether
distinct brain regions are activated during production of regular and
non‐regular verbs.
References
Aitchison, J. (1997). Words in the mind: An introduction to the mental lexicon.
Oxford: Blackwell.
Alario, F. X. & Caramazza, A. (2002). The production of determiners: Evidence
from French. Cognition, 82, 179–223.
Alegre, M. & Gordon, P. (1999). Frequency effects and the representational status
of regular inflection. Journal of Memory and Language, 40, 41–61.
Aronoff, M. (1976). Word formation in Generative Grammar. Cambridge,
Massachusetts: MIT Press.
Aronoff, M. (2000). Morphology between lexicon and grammar. In: G. E. Booij, C.
Lehmann, J. Mugdan, & S. Skopeteas (Eds.), Morphology. An international
handbook on inflection and word formation (pp. 344–349). Berlin: De Gruyter.
Baayen, H., Piepenbrock, R. & van Rijn, H. (1993). The CELEX lexical database
(CD‐ROM). University of Pennsylvania: Linguistic Data Consortium.
Baayen, H. R., Dijkstra, T. & Schreuder, R. (1997). Singulars and Plurals in Dutch:
Evidence for a Parallel Dual‐Route Model. Journal of Memory and Language, 37,
94–117.
Beck, M. (1997). Regular verbs, past tense and frequency: Tracking down a
potential source for NS/NNS competence differences. Second Language
Research, 13(2), 93–115.
Beretta, A., Campbell, C., Carr, T. H., Huang, J., Schmitt, L. M., Christianson, K. &
Cao, Y. (2003). An ER‐fMRI investigation of morphological inflection in German
reveals that the brain makes a distinction between regular and irregular forms.
Brain and Language, 85, 67–92.
Beringer, J. (1995). ERTS: Experimental run time system. Frankfurt am Main:
Berisoft Cooperation.
Berko, J. (1958). The child's learning of English morphology. Word, 14, 150–177.
Bickel, B. & Nichols, J. (2007). Inflectional morphology. In: T. Shopen (Ed.),
Language typology and syntactic description (2nd ed.). Cambridge: Cambridge
University Press.
Bock, K. (2004). Psycholinguistically speaking: Some matters of meaning, marking,
and morphing. The Psychology of Learning and Motivation, 44, 109–144.
Bordag, D. & Pechmann, T. (2009). Externality, internality, and (in)dispensability of
grammatical features in speech production: Evidence from Czech declension and
conjugation. Journal of Experimental Psychology: Learning, Memory and
Cognition, 35(2), 446–465.
157
158 REFERENCES
Branigan, H. P., Pickering, M. J., Liversedge, S. P., Stewart, A. J. & Urbach, T. P.

(1995). Syntactic Priming: Investigating the Mental Representation of Language.
Journal of Psycholinguistic Research, 24(6), 489–506.
Butterworth, B. (1983). Lexical representations. In: B. Butterworth (Ed.), Language
Production. New York: Academic Press.
Bybee, J. L. (1985). Morphology. A study of the relation between meaning and
form. Amsterdam; Philadelphia: John Benjamins.
Bybee, J. L. (1988). Morphology as lexical organisation. In: M. Hammond & M.
Noonan (Eds.), Theoretical morphology. San Diego, CA: Academic Press.
Bybee, J. L. & Moder, C. L. (1983). Morphological classes as natural categories.
Language, 59, 251–270.
Caramazza, A. (1997). How many levels of processing are there in lexical access?
Cognitive Neuropsychology, 14, 177–208.
Caramazza, A. & Costa, A. (2000). The semantic interference effect in the picture‐
word interference paradigm: does the response set matter? Cognition, 75, B51‐
B64.
Caramazza, A. & Hillis, A. E. (1991). Lexical organization of nouns and verbs in the
brain. Nature, 349(6312), 788–790.
Caramazza, A., Laudanna, A. & Romani, C. (1988). Lexical access and inflectional
morphology. Cognition, 28, 297–332.
Caramazza, A., Miceli, G., Silveri, M. C. & Laudanna, A. (1985). Reading
mechanisms and the organization of the lexicon: evidence from acquired
dyslexia. Cognitive Neuropsychology, 2, 81–114.
Caramazza, A. & Miozzo, M. (1997). The relation between syntactic and
phonological knowledge in lexical access: Evidence from the “Tip of the
Tongue” phenomenon. Cognition, 64, 309–349.
Carstairs, A. (1987). Allomorphy in inflection. London: Croom Helm.
Carstairs‐McCarthy, A. D. (2000). Inflection classes. In: G. E. Booij, C. Lehmann, J.
Mugdan, & S. Skopeteas (Eds.), Morphology. An international handbook on
inflection and word formation (pp. 630–637). Berlin: De Gruyter.
Chomsky, N. (1995). The Minimalist Program. Cambridge, Massachusetts: MIT
Press.
Chomsky, N. & Halle, M. (1968). The sound pattern of English. New York: Harper
and Row.
Clahsen, H. (1996). The representation of German participles in the German mental
lexicon: evidence for the dual‐mechanism model. Yearbook of Morphology, 73–
95.
Clahsen, H. (1999). Lexical entries and rules of language: A multi‐disciplinary
study of German inflection. Behavioral and Brain Sciences, 22, 991–1060.
Clahsen, H., Aveledo, F. & Roca, I. (2002). The development of regular and
irregular verb inflection in Spanish child language. Journal of Child Language,
29(3), 591–622.
159
Clahsen, H., Eisenbeiss, S., Hadler, M. & Sonnenstuhl, I. (2001). The mental
representation of inflected words: an experimental study of adjectives and verbs
in German. Language, 77(3), 510–543.
Clahsen, H., Eisenbeiss, S. & Sonnenstuhl, I. (1997). Morphological structure and
the processing of inflected words. Theoretical Linguistics, 23, 201–249.
Clahsen, H., Hadler, M. & Weyerts, H. (2004). Speeded production of inflected
words in children and adults. Journal of Child Language, 31, 683–712.
Clahsen, H., Rothweiler, M., Woest, A. & Marcus, G. F. (1992). Regular and
irregular inflection in the acquisition of German noun plurals. Cognition, 45,
225–255.
Collins, R. & Loftus, E. F. (1975). A spreading activation theory of semantic
processing. Psychological Review, 82, 407–428.
Coltheart, M., Curtis, B., Atkins, P. & Haller, M. (1993). Models of reading aloud:
Dual‐route and parallel‐distributed‐processing approaches. Psychological
Review, 100(4), 589–608.
Coltheart, M. (2005). What has functional neuroimaging told us about the mind (so
far)? Position paper presented to the European Cognitive Neuropsychology
Workshop, Bressanone.
Comrie, B. (1987). Tense. Cambridge: Cambridge University Press.
Comrie, B. (1989). Language universals and linguistic typology (2nd edition).
Chicago: University of Chicago.
Cutting, J. C. (1998). The production and comprehension lexicons: What is shared
and what is not. Dissertation Abstracts International: Section B: The Sciences and
Engineering, Vol 58(10‐B), pp. 5663.
Daugherty, K. & Seidenberg, M. S. (1994). Beyond rules and exceptions: A
connectionist approach to inflectional morphology. In: S. D. Lima, R. Corrigan, &
G. Iverson (Eds.), The reality of linguistic rules. Amsterdam: John Benjamins.
Davidoff, J. & Masterson, J. (1996). The development of picture naming:
Differences between verbs and nouns. Journal of Neurolinguistics, 9, 69–83.
Dell, G. S. (1986). A spreading activation model of retrieval in sentence production.
Psychological Review, 93, 283–321.
Dell, G. S. & Reich, P. A. (1981). Stages in sentence production: An analysis of
speech error data. Journal of Verbal Learning and Verbal Behavior, 20, 611–629.
Dohmes, P., Zwitserlood, P. & Bölte, J. (2004). The impact of semantic transparency
of morphologically complex words on picture naming. Brain and Language, 90,
203–212.
Donders, F. C. (1868, 1969). On the speed of mental processes. Acta Psychologica,
30, 412–431.
Druks, J. (2002). Verbs and Nouns – a review of the literature. Journal of
Neurolinguistics, 15, 289–315.
Dudenredaktion (1998). Duden ‐ Grammatik der deutschen Gegenwartssprache.
Mannheim: Dudenverl.
Eisenberg, P. (1994). Grundriß der deutschen Grammatik (3rd ed.). Stuttgart [u.a.]:
Metzler.
160 REFERENCES
Elman, J. L. (2004). An alternative view of the mental lexicon. Trends in Cognitive

Science, 8(7), 301–306.
Faroqi‐Shah, Y. (2007). Are regular and irregular verbs dissociated in non‐fluent
aphasia? A meta‐analysis. Brain Research Bulletin, 74, 1–13.
Faroqi‐Shah, Y. & Thompson, C. K. (2007). Verb inflections in agrammatic aphasia:
encoding of tense features. Journal of Memory and Language, 56, 129–151.
Finkbeiner, M. & Caramazza, A. (2006). Now you see it, now you don’t: On turning
semantic interference into facilitation in a Stroop‐like task. Cortex, 42, 790–796.
Forster, K. I. & Chambers, S. M. (1973). Lexical access and naming time. Journal of
Verbal Learning and Verbal Behavior, 12, 627–635.
Friederici, A. D. (1995). The time course of syntactic activation during language
processing: a model based on neuropsychological and neurophysiological data.
Brain and Language, 50, 259–281.
Fromkin, V. A. (1973). Speech errors as linguistic evidence. The Hague: Mouton.
Garrett, M. F. G. (1975). The analysis of sentence production. In: G. H. Bower (Ed.),
The psychology of learning and motivation. Academic Press.
Garrett, M. F. G. (1982). Production of speech: Observation from normal and
pathological language use. In: A. W. Ellis (Ed.), Normality and pathology in
cognitive functions (pp. 19–76). London: Academic Press.
Gibbs, R. W. (1980). Spilling the beans on understanding and memory for idioms
conversation. Memory and Cognition, 8(2), 149–156.
Gordon, R. G. (2005). Ethnologue: Languages of the World, Fifteenth edition.
Dallas: SIL International.
Hahne, A. & Friederici, A. D. (1999). Electrophysiological evidence for two steps in
syntactic analysis: early automatic and late controlled processes. Journal of
Cognitive Neuroscience, 11, 194–205.
Halle, M. & Marantz, A. (1993). Distributed Morphology and the Pieces of
Inflection. The View from Building, 20, 111–176.
Halle, M. & Mohanan, K. (1985). Segmental phonology of modern English.
Linguistic Inquiry, 16, 57–166.
Heidolph, K. E., Fläming, W. & Motsch, W. (1981). Grundzüge einer deutschen
Grammatik. Berlin: Akademie Verlag.
Henning, M. (2000). Tempus und Temporalität in geschriebenen und gesprochenen
Texten. Linguistische Arbeiten: Vol. 421. Tübingen.
Humphreys, G. W., Riddock, M. J. & Quinlan, P. T. (1988). Cascade processes in
picture identification. Cognitive Neuropsychology, 5, 67–104.
Jaeger, J. J., Lockwood, A. H., Kemmerer, D. L., van Valin, R. D., Murphy, B. W. &
Khalak, H. G. (1996). A positron emission tomographic study of regular and
irregular verb morphology in English. Language, 72, 451–497.
Janssen, D. P. (1999). Producing past and plural inflections. Unpublished doctoral
dissertation. Nijmegen: Nijmegen University.
Jescheniak, J. D. & Levelt, W. J. M. (1994). Word frequency effects in speech
production: Retrieval of syntactic information and of phonological form. Journal
of Experimental Psychology: Learning, Memory and Cognition, 20(4), 824–843.
161
Joanisse, M. F. & Seidenberg, M. S. (1999). Impairments in verb morphology after

brain injury: A connectionist model. Proceedings of the National Academy of
Sciences of the United States of America, 7592–7597.
Kauschke, C. (2007). Erwerb und Verarbeitung von Nomen und Verben. Tübingen:
Niemeyer.
Kempen, G. & Hoenkamp, E. (1987). An incremental procedural grammar for
sentence formulation. Cognitive Science, 11, 201–258.
Kempen, G. & Huijbers, P. (1983). The lexicalisation process in sentence production
and naming. Indirect election of words. Cognition, 14, 185–209.
Kempley, S. & Morton, J. (1982). The effect of priming with regularly and
irregularly inflected words in auditory word recognition. British Journal of
Psychology, 73, 441–454.
Kunkel‐Razum, K. (2006). Duden ‐ die Grammatik (Nach den Regeln der neuen dt.
Rechtschreibung 2006 überarb. Neudr. der 7., völlig neu erarb. und erw. Aufl.).
Der Duden: Bd. 4. Mannheim: Dudenverl.
La Heij, W., Mak, P., Sander, J. & Willeboordse, E. (1998). The gender congruency
effect in picture‐word tasks. Psychological Research, 61, 209–219.
La Heij, W. & van den Hof, E. (1995). Picture‐word interference increases with
target‐set size. Psychological Research, 58(2), 119–133.
Laiacona, M. & Caramazza, A. (2004). The noun/verb dissociation in language
production. Varieties of causes. Cognitive Neuropsychology, 21(2‐4), 103–123.
Levelt, W. J. M. (1989). Speaking. From intention to articulation. Cambridge,
Massachusetts: MIT Press.
Levelt, W. J. M. (1991). Die konnektionistische Mode. Sprache und Kognition, 10(2),
61–71.
Levelt, W. J. M., Roelofs, A. & Meyer, A. S. (1999). A theory of lexical access in
speech production. Behavioral and Brain Sciences, 22, 1–38.
Levelt, W. J. M., Schriefers, H., Vorberg, D., Meyer, A. S., Pechmann, T. & Having,
J. (1991). The time course of lexical access in speech production: A study of
picture naming. Psychological Review, 98, 122–142.
Ling, C. & Marinov, M. (1993). Answering the connectionist challenge: A symbolic
model of learning the past tenses of English verbs. Cognition, 49(3), 235–290.
Logothetis, N. K. (2008). What we can do and what we cannot do with fMRI.
Nature, 453, 869–878.
Lück, M., Hahne, A. & Clahsen, H. (2006). Brain potentials to morphologically
complex words during listening. Brain Research, 1077(1), 144–152.
Lukatela, G., Gligorijevic, B., Kostic, A. & Turvey, M. T. (1980). Representation of
inflected nouns in the internal lexicon. Memory and Cognition, 8, 415–423.
MacWhinney, B. & Leinbach, J. (1991). Implementations are not conceptualisations.
Revising the verb learning model. Cognition, 40, 121–157.
Mahon, B. Z., Costa, A., Peterson, R. R., Vargas, K. A. & Caramazza, A. (2007).
Lexical selection is not by competition: A reinterpretation of semantic
interference and facilitation effects in the picture–word interference paradigm.
162 REFERENCES
Journal of Experimental Psychology: Learning, Memory and Cognition, 33(3),

503–535.
Marcus, G. F., Brinkmann, U., Clahsen, H., Wiese, R. & Pinker, S. (1995). German
inflection: The exception that proves the rule. Cognitive Psychology, 29, 189–256.
Marcus, G. F., Pinker, S., Ullman, M. T., Hollander, M., Rosen, T. J. & Xu, F. (1992).
Overregularization in language acquisition. Monographs of the Society for
Research in Child Development, 57, 4, Serial No. 228.
Marslen‐Wilson, W. D. & Tyler, L. K. (1997). Dissociating types of mental
computation. Nature, 387, 592–594.
Marslen‐Wilson, W. D. & Tyler, L. K. (1998). Rules, representations, and the
English past tense. Trends in Cognitive Science, 2(11), 428–435.
Masterson, J. & Druks, J. (1998). Description of a set of 164 nouns and 102 verbs
matched for printed word and frequency, familiarity and age‐of‐acquisition.
Journal of Neurolinguistics, 11(4), 331–354.
McClelland, J. L. & Rumelhart, D. (1986). On learning the past tenses of English
verbs. In: D. Rumelhart, J. L. McClelland, & the PDP research group (Eds.),
Parallel distributed processing: Explorations in the microstructure of cognition.
MIT Press.
Meunier, F. & Marslen‐Wilson, W. D. (2004). Regularity and irregularity in French
verbal inflection. Language and Cognitive Processes, 19(4), 561–580.
Miozzo, M. (2003). On the processing of regular and irregular forms of verbs and
nouns: evidence from neuropsychology. Cognition, 87, 101–127.
Miozzo, M. & Caramazza, A. (1997). Retrieval of lexical‐syntactic features in tip‐of‐
the‐tongue‐states. Journal of Experimental Psychology: Learning, Memory and
Cognition, 23, 1410–1423.
Miozzo, M. & Caramazza, A. (1999). The selection of determiners in noun phrase
production. Journal of Experimental Psychology: Learning, Memory and
Cognition, 25, 907–922.
Miozzo, M. & Caramazza, A. (2003). When more is less: A counterintuitive effect of
distractor frequency in the picture‐word interference paradigm. Journal of
Experimental Psychology: General, 132(2), 228–252.
Münte, T. F., Rodriguez‐Fornells, A. & Kutas, M. (1999). One, two, or many
mechanisms? The brain's processing of complex words. Behavioral and Brain
Sciences, 22(6), 1031–1032.
Newman, A. J., Ullman, M. T., Pancheva, R., Waligura, D. L. & Neville, H. J. (2007).
An ERP study of regular and irregular past tense inflection. NeuroImage, 34,
435–445.
Oldfield, R. C. & Wingfield, A. (1965). Response latencies in naming objects.
Quarterly Journal of Experimental Psychology, 17, 273–281.
Orsolini, M. & Marslen‐Wilson, W. D. (1997). Universals in morphological
representation: Evidence from Italian. Language and Cognitive Processes, 12(1),
1–47.
Osterhout, L. & Mobley, L. A. (1995). Event‐related brain potentials elicited by
failure to agree. Journal of Memory and Language, 34, 739–773.
163
Page, M. P. A. (2006). What can't functional neuroimaging tell the cognitive

psychologist? Cortex, 42(3), 428–443.
Patterson, K., Lambon Ralph, M. A., Hodges, J. R. & McClelland, J. L. (2001).
Deficits in irregular past‐tense verb morphology associated with degraded
semantic knowledge. Neuropsychologia, 39, 709–724.
Pechmann, T., Garrett, M. F. G. & Zerbst, D. (2004). The time course of recovery for
grammatical category information during lexical processing for syntactic
construction. Journal of Experimental Psychology: Learning, Memory and
Cognition, 30(3), 723–728.
Pechmann, T., Reetz, H. & Zerbst, D. (1989). Kritik einer Messmethode: zur
Ungenauigkeit von vioce‐key Messungen. Sprache & Kognition, 8, 65–71.
Pechmann, T. & Schriefers, H. (unpublished). The role of identical distractors in
picture‐word interference experiments.
Pechmann, T. & Zerbst, D. (2002). The activation of word class information during
speech production. Journal of Experimental Psychology: Learning, Memory and
Cognition, 28(1), 233–243.
Penke, M. (2006). Flexion im mentalen Lexikon. Linguistische Arbeiten: Vol. 503.
Tübingen: Niemeyer (Univ., Habil‐Schr.‐‐Düsseldorf).
Penke, M. & Krause, M. (2002). German noun plurals: A challenge to the Dual‐
Mechanism Model. Brain and Language, 81, 303–311.
Penke, M., Weyerts, H., Gross, M., Zander, E., Münte, T. F. & Clahsen, H. (1997).
How the brain processes complex words: An event‐related potential study of
German verb inflections. Cognitive Brain Research, 6(1), 37–52.
Peterson, R. R. & Savoy, P. (1998). Lexical selection and phonological encoding
during language production: Evidence for cascaded processing. Journal of
Experimental Psychology: Learning, Memory and Cognition, 24, 539–557.
Pinker, S. (1991). Rules of language. Science, 253, 530–535.
Pinker, S. (1994). The Language Instinct. New York: William Morrow.
Pinker, S. (1997). Words and rules in the human brain. Nature, 387, 547–548.
Pinker, S. (1999). Words and rules: The ingredients of language. New York: Harper
Collins.
Pinker, S. & Prince, A. (1994). Regular and irregular Morphology and the
psychological status of rules of Grammar. In: S. D. Lima, R. Corrigan, & G.
Iverson (Eds.), The reality of linguistic rules. Amsterdam: John Benjamins.
Pinker, S. & Ullman, M. T. (2002). The past and future of the past tense. Trends in
Cognitive Science, 6, 456–463.
Plunkett, K. & Marchman, V. (1993). From rote learning to system building:
Acquiring verb morphology in children and connectionist nets. Cognition, 48(1),
21–69.
Prasada, S., Pinker, S. & Snyder, W. (1990). Some evidence that irregular forms are
retrieved from memory but regular forms are rule‐genrated. Paper presented to
the Annual Meeting of the Psychonomic Society, New Orleans, November 23‐25.
Prince, A. & Pinker, S. (1988). Rules and connections in human language. Trends in
Neurosciences, 11(5), 195–202.
164 REFERENCES
Prinz, W. (1997). Perception and action planning. European Journal of Cognitive

Psychology, 9(2), 129–154.
Pustejovsky, J. (1996). The Generative Lexicon: MIT Press.
Ramscar, M. (2002). The role of meaning in flection: Why the past tense does not
require a rule. Cognitive Psychology, 45(1), 45–94.
Rapp, B. & Golrick, M. (2006). Speaking words: Contributions of cognitive
neuropsychological research. Cognitive Neuropsychology, 23(1), 39–73.
Rastle, K., Davis, M. H., Marslen‐Wilson, W. D. & Tyler, L. K. (2000).
Morphological and semantic effects in visual word recognition: A time‐course
study. Language and Cognitive Processes, 15, 507–537.
Reichenbach, H. (1947). Elements of symbolic logic. New York: Dover.
Reicher, G. M. (1969). Perceptual recognition as a function of meaningfulness of
stimulus matrial, 81,. Journal of Experimental Psychology, 81(2), 275–280.
Rodriguez, P., Wiles, J. & Elman, J. L. (1999). A recurrent neural network that
learns to count. Connection Science, 11(1), 5–40.
Roelofs, A. (1992). A spreading activation theory of lemma retrieval in speaking.
Cognition, 42, 107–142.
Roelofs, A. (1993). Testing a non decompositional theory of lemma retrieval in
speaking: Retrieval of verbs. Cognition, 47, 59–87.
Roelofs, A. (1997). The weaver model of word‐form encoding in speech
production. Cognition, 64, 249–284.
Rumelhart, D. & J. L. McClelland & the PDP research group (Eds.) (1986). Parallel
distributed processing: Explorations in the microstructure of cognition: MIT
Press.
Sahin, N., Pinker, S. & Halgren, E. (2006). Abstract Grammatical Processing of
Nouns and Verbs in Broca's Area: Evidence from fMRI. Cortex, 42, 540–562.
Schiller, N. O. & Caramazza, A. (2002). The selection of grammatical features in
word production: The case of plural nouns in German. Brain and Language, 81,
342–357.
Schiller, N. O. & Caramazza, A. (2003). Grammatical feature selection in noun
phrase production: Evidence from German and Dutch. Journal of Memory and
Language, 48, 169–194.
Schmidt, W., Langner Helmut & Wolf, N. R. (1996). Geschichte der deutschen
Sprache. Stuttgart: Hirzel.
Schriefers, H. (1993). Syntactic processes in the production of noun phrases. Journal
of Experimental Psychology: Learning, Memory and Cognition, 19, 841–850.
Schriefers, H. & Jescheniak, J. D. (1999). Representation and processing of
grammatical gender in language production: A review. Journal of
Psycholinguistic Research, 28(6), 575–600.
Schriefers, H., Meyer, A. S. & Levelt, W. J. M. (1990). Exploring the time course of
lexical access in language production: Picture‐word interference studies. Journal
of Memory and Language, 29, 86–102.
165
Schriefers, H. & Teruel, E. (2000). Grammatical gender in noun phrase production:

The gender interference effect in German. Journal of Experimental Psychology:
Learning, Memory and Cognition, 26, 1368–1377.
Seidenberg, M. S. (1992). Connectionism without tears. In: S. Davis (Ed.),
Connectionism: Theory and practice (pp. 84–122). New York, NY: Oxford
University Press.
Seidenberg, M. S. & Bruck, M. (1990). Consistency effects in the generation of past
tense morphology. Paper presented to the Annual Meeting of the Psychonomic
Society, New Orleans, November 23‐25.
Shattuck‐Hufnagel, S. (1979). Speech errors as evidence for a serial‐ordering
mechanism in sentence production. In: W. E. Cooper & E. C. T. Walker (Eds.),
Sentence processing: Psycholinguistic studies dedicated to Merrill Garrett.
Erlbaum.
Simmons‐Mackie, N. N. & Damico, J. S. (1997). Reformulating the definition of
compensatory strategies in aphasia. Aphasiology, 11(8), 761–781.
Siri, S., Tettamanti, M., Cappa, S. F., Della‐Rosa, P., Saccuman, C., Scifo, P. &
Vigliocco, G. (2008). The neural substrate of naming events: Effects of processing
demands but not of grammatical class. Cerebral Cortex, 18(1), 171–177.
Skousen, R. (1992). Analogy and structure. Dordrecht: Kluwer Academic.
Smith, E. E., Shoben, E. J. & Rips, L. J. (1974). Structure and processes in semantic
memory. A featural model for semantic decisions. Psychological Review, 81,
214–241.
Sonnenstuhl, I., Eisenbeiss, S. & Clahsen, H. (1999). Morphological priming in the
German mental lexicon. Cognition, 72, 203–236.
Sonnenstuhl, I. & Huth, A. (2002). Processing and Representation of German n‐
plurals: A dual mechanism approach. Brain and Language, 81, 276–290.
Spencer, A. & Zwicky, A. (1998). The handbook of morphology. Oxford: Blackwell.
Stanners, R. F., Neiser, J. J., Hernon, W. P. & Hall, R. (1979). Memory representation
for morphologically related words. Journal of Verbal Learning and Verbal
Behavior, 18(4), 399–412.
Starreveld, P. A. & La Heij, W. (1995). Semantic interference, orthographic
facilitation and their interaction in naming tasks. Journal of Experimental
Psychology: Learning, Memory and Cognition, 21, 686–698.
Starreveld, P. A. & La Heij, W. (1996). Time‐course analysis of semantic and
orthographic context effects in picture naming. Journal of Experimental
Psychology: Learning, Memory and Cognition, 22, 896–918.
Stroop, J. R. (1935). Studies of interference in serial verbal reactions. Journal of
Experimental Psychology, 18, 643–662.
Swinney, D. & Cutler, A. (1979). The access and processing of idiomatic
expressions. Journal of Verbal Learning and Verbal Behavior, 18(5), 523–534.
Szekely, A., DÀmico, S., Devescovi, A., Federmeier, K., Herron, D., Iyer, G.,
Jacobsen, T. & Bates, E. (2005). Timed action and object naming. Cortex, 41(1), 7–
26.
166 REFERENCES
Tabossi, P., Collina, S. & Sanz, M. (2002). The retrieval of syntactic and semantic
information in the production of verbs. Brain and Language, 81, 264–275.
Taft, M. (1981). Prefix stripping revisited. Journal of Verbal Learning and Verbal
Behavior, 20, 289–297.
Taft, M. & Forster, K. I. (1975). Lexical storage and retrieval of prefixed words.
Journal of Verbal Learning and Verbal Behavior, 14, 638–647.
Tylor, L. K., Bright, P., Fletcher, P. & Stamatakis, E. A. (2004). Neural processing of
nouns and verbs: the role of inflectional morphology. Neuropsychologia, 42,
512–523.
Ullman, M. T. (2001a). A neurocognitive perspective on language: the
declarative/procedural model. Nature Neuroscience, 2, 717–726.
Ullman, M. T. (2001b). The declarative procedural model of lexicon and grammar.
Journal of Psycholinguistic Research, 30, 37–69.
Ullman, M. T., Corkin, S., Coppola, M. & Hickok, G. (1997). A neural dissociation
within language: evidence that the mental dictionary is part of declarative
memory and that grammatical rules are processed by the procedural system.
Journal of Cognitive Neuroscience, 9(2), 266–276.
van Berkum, J. J. A. (1997). Syntactic processes in speech production: The retrieval
von grammatical gender. Cognition, 64, 115–152.
Vigliocco, G., Vinson, D. P., Damian, M. F. & Levelt, W. J. M. (2002). Semantic
distance effects on object and action naming. Cognition, 85, 61–69.
Vigliocco, G., Vinson, D. P., Lewis, W. & Garrett, M. F. G. (2004). Representing the
meanings of object and action words: The featural and unitary semantic space
hypothesis. Cognitive Psychology, 48, 422–488.
Vigliocco, G., Vinson, D. P. & Siri, S. (2005). Semantic similarity and grammatical
class in naming actions. Cognition, 94, B91_B100.
Whaley, C. P. (1978). Word‐non‐word classification time. Journal of Verbal
Learning and Verbal Behavior, 17, 143–154.
Wiese, B. (in prep.). Form and Function of verbal ablaut in Modern standard
German. In: R. Sackmann (Ed.), Studies in Integrational Linguistics 1.
Amsterdam; Philadelphia: Benjamins.
Wunderlich, D. (1996). Minimalist morphology. The role of paradigms. Yearbook
of Morphology 1995, 93–114.
Wurzel, W. U. (1984). Flexionsmorphologie und Natürlichkeit. Ein Beitrag zur
morphologischen Theoriebildung. Berlin: Akademie Verlag.
Zwitserlood, P. (1994). Access to phonological‐form representations in language
comprehension and production. In: C. L. Clifton, L. Frazier, & K. Rayner (Eds.),
Perspectives on sentence processing (pp. 83–106). Hillsdale, NJ: Lawrence
Erlbaum Associates, Inc.
Zwitserlood, P., Bölte, J. & Dohmes, P. (2002). Where and how morphologically
complex words interplay with naming pictures. Brain and Language, 81(1‐3),
358–367.
Appendix
Stimuli for Experiment 1, 2 and 3

regular verbs – 3rd present, 3rd past
angeln [to fish] – er angelt, er angelte

beten [to pray] – er betet, er betete
betteln [to beg] – er bettelt, er bettelte
bluten [to bleed] – er blutet, er blutete
bohren [to drill] – er bohrt, er bohrte
klopfen [to knock] – er klopft, er klopfte
knien [to kneel] – er kniet, er kniete
kochen [to cook] – er kocht, er kochte
lachen [to laugh] ‐ er lacht, er lachte
läuten [to ring] – er läutet, er läutete
niesen [to sneeze] – er niest, er nieste
schaukeln [to swing] – er schaukelt, er schaukelte
tanzen [to dance] – er tanzt, er tanzte
weben [to weave] – er webt, er webte
weinen [to cry] – er weint, er weinte
winken [to wave] – er winkt, er winkte
non-regular verbs – 3rd present, 3rd past
beißen [to bite] – er beißt, er biss

essen [to eat] – er isst, er aß
fahren [to drive] – er fährt, er fuhr
fangen [to snatch] – er fängt, er fing
fliegen [to fly] – er fliegt, er flog
gießen [to water] – er gießt, er goss
graben [to dig] – er gräbt, er grub
lesen [to read] – er liest, er las
rennen [to run] – er rennt, er rannte
schieben [to push] – er schiebt, er schob
schlafen [to sleep] – er schläft, er schlief
schwimmen [to swim] – er schwimmt, er schwamm
167
168 APPENDIX
singen [to sing] ‐ er singt, er sang

springen [to jump] – er springt, er sprang
tragen [to carry] – er trägt, er trug
trinken [to drink] – er trinkt, er trank
Stimuli for Experiment 4 and 5

regular verbs – 3rd present, 3rd past
bauen [to build] – er baut, er baute

betteln [to beg] – er bettelt, er bettelte
bohren [to drill] – er bohrt, er bohrte
gähnen [to yawn] – er gähnt, er gähnte
klopfen [to knock] – er klopft, er klopfte
küssen [to kiss] – er küsst, er küsste
lachen [to laugh] ‐ er lacht, er lachte
tanzen [to dance] – er tanzt, er tanzte
tauchen [to dive] – er taucht, er tauchte
hybrid verbs – 3rd present, 3rd past
beißen [to bite] – er beißt, er biss

biegen [to bend] – er biegt, er bog
gießen [to water] – er gießt, er goss
greifen [to grasp] – er greift, er griff
kneifen [to pinch] – er kneift, er kniff
kriechen [to crawl] – er kreicht, er kroch
pfeifen [to whistle] – er pfeift, er pfiff
trinken[to drink] – er trinkt, er trank
ziehen [to pull] – er zieht, er zog
irregular verbs – 3rd present, 3rd past
blasen [to blow] – er bäst, er bließ

braten [to fry] – er brät, er briet
brechen [to vomit] – er bricht, er brach
dreschen [to thresh] – er drischt, er drosch
essen [to eat] – er isst, er aß
graben [to dig] – er gräbt, er grub
lesen [to read] – er liest, er las
tragen [to carry] – er trägt, er trug
treten [to kick] – er tritt, er trat
The incredible productivity and creativity of language depends on two
fundamental resources: a mental lexicon and a mental grammar. Rules
of grammar enable us to produce and understand complex phrases we
have not encountered before and at the same time constrain the compu-
tation of complex expressions. The concepts of the mental lexicon and
mental grammar have been thoroughly tested by comparing the use of
regular versus non-regular word forms. Regular verbs (e.g. walk-walked)
are computed using a suffixation rule in a neural system for grammatical
processing; non-regular verbs (run-ran) are retrieved from associative
memory. The role of regularity has only been explored for the past tense,
where regularity is overtly visible.
To explore the representation and encoding of regularity as well as the in-
flectional processes involved in the production of regular and non-regular
verbs, this dissertation investigated three groups of German verbs: regu-
lar, irregular and hybrid verbs. Hybrid verbs in German have completely
regular conjugation in the present tense and irregular conjugation in the
past tense. Articulation latencies were measured while participants named
pictures of actions, producing the 3rd person singular of regular, hybrid,
and irregular verbs in present and past tense. Studying the production
of German verbs in past and present tense, this dissertation explored the
complexity of lexical entries as a decisive factor in the production of verbs.
ISSN 1869-3822
ISBN 978-3-86956-061-8

Regular & Irregular Verbs PDF

Uploaded by

Copyright:

Available Formats

Regular & Irregular Verbs PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Regular & Irregular Verbs PDF

Uploaded by

Copyright:

Available Formats

Universität Potsdam

Production of regular and

Spektrum Patholinguistik – Schriften | 2

Production of regular and non-regular verbs

Universitätsverlag Potsdam 2010

Universitätsverlag Potsdam, Am Neuen Palais 10, 14469 Potsdam

Die Schriftenreihe Spektrum Patholinguistik – Schriften wird herausgegeben

Das Manuskript ist urheberrechtlich geschützt.

Satz: Martin Anselm Meyerhoff

Zugl.: Potsdam, Univ., Diss., 2010

ISSN (print) 1869‐3822

Zugleich online veröffentlicht

First and foremost, I am grateful to Prof. Dr. Thomas Pechmann

The Graduiertenkolleg Universalität und Diversität at the

I am indebted to Dr. habil. Denisa Bordag for her constant and

I would like to thank my friends Lars Meyer, Judith Heide, Tyko

smart comments on a first version of this thesis and for

Most important of all: I count myself truly fortunate for having

Finally, the greatest debts of gratitude by far are to my parents

List of Tables........................................................................ vii

1 Regular and non-regular inflection................................... 5

1.1 Inflectional categories .......................................................... 5

2 Approaches to regular and non-regular inflection ............19

2.1 Articulation latencies of regular and non-regular verbs ....... 19

2.3.2 Psycholinguistic evidence for a regular/non-regular

3 Psycholinguistic models of language production ..............47

3.1 Lexical access and lexical selection..................................... 47

4 Representation and processing of grammatical features ..71

4.1 Representation of linguistic information in the mental

6 The empirical stance......................................................87

6.1 Why and how regularity might be represented ................... 87

7.1 Experiment 1 – Present tense ............................................ 95

Stimuli for Experiment 1, 2 and 3................................................. 167

TABLE 16. ACCURACY PROPORTIONS, BY REGULARITY, TENSE AND

Moreover, rules apply on abstract symbolic categories rather than

The comparison by Steven Pinker of regular verbs being the fruit

1 The term regularity includes non‐regularity. It is a superordinate concept and

the different morphological inflectional processes assumed and

1.1 Inflectional categories

Inflection is one way languages express grammatical relations.

2 A non‐inherent inflectional category of verbal inflection is for example subject‐

Nichols, 2007). Typologically, different languages may also specify

(1) ich spiel‐e, er spiel –t, er spiel‐te [I play, he play‐s, he play‐ed]

(3) a. he sing-s/we sing

“The relevant grammatical environment can be either syntactic or

By contrast to syntactically driven inflectional processes, the

1.2 Paradigms and classes

its grammatical function. For German, see example (4) for

(4) a. ich lach‐e, du lach‐st, er lach‐t…

The person/number suffixes for conjugating lachen and gehen are

and singen [to sing] belong to the non‐regular paradigm but to

1.3 Language typology

Anglo‐American research constitutes a large part of the

vowel alternation to indicate past tense. These are known as

1.3.1 German verbal inflectional system

The German verb system is indeed organised in a greater diversity

(5) ich spiel‐e, er spiel –t, er spiel‐te [I play, he play‐s, he play‐ed]

Hybrid verbs unify regular and irregular forms in one paradigm.

3 Weak and strong is a synonym terminology for regular and non‐regular.