ELSEVIER Journal of Pragmatics 32 (2000) 1247-1252

www.elsevier.nldocate/pragma

Book review

Walter Kintseh, Comprehension: A paradigm for cognition. Cambridge: Cam-

bridge University Press, 1998. xvi + 461 pp. $69.95 (hb.), $27.95 (pb.).

Reviewed by Arthur Graesser and Shannon Whitten, Department of Psychology,

Campus Box 526400, The University of Memphis, Memphis, TN 38152, USA. E-
mail: a-graesser@ memphis.edu I

This is the book to read for anyone who wants to become up-to-date in the field
of discourse psychology. What is discourse psychology? Discourse psychologists
investigate the psychological mechanisms that underlie the comprehension and pro-
duction of discourse (Graesser et al., 1997). These psychological mechanisms
include a set of processes that recruit knowledge structures in long-term memory and
that result in a representation of the text or utterances. When adults read a story, for
example, they perceive the words, group the words into constituents, activate rele-
vant world knowledge, construct inferences, and settle on coherent meaning repre-
sentations. Each of these component processes takes time. Claims about the process-
ing stream are tested in experiments that collect reading times for words or sentences
from a sample of readers, or that measure the extent to which particular ideas are
activated in the mind of the reader (as reflected in word naming latencies or lexical
decision latencies). Each level of representation is coded in some fashion. Claims
about the representation are tested in experiments that collect memory protocols,
think aloud protocols, truth verification judgments, and many other tasks. A good
theory in discourse psychology produces output that has a close fit to the data col-
lected in these experiments. A better theory also is articulated in a mathematical
model and is compatible with general theories of cognition in the cognitive sciences.
Kintsch's theory is currently the most popular and comprehensive theory that exists
in the field of discourse psychology.
Discourse psychology is an interdisciplinary field in the sense that efforts are
made to incorporate the empirical findings and theoretical insights from diverse dis-
ciplines in the cognitive sciences. However, it is important to acknowledge that the

Partial support for the preparation of this review came from a National Science Foundation grant
(SBR 9720314).

0378-2166/00/$ - see front matter © 2000 Elsevier Science B.V. All rights reserved.
PI: S 0 3 7 8 - 2 1 6 6 ( 9 9 ) 0 0 0 9 0 - 9
 1248 Book review / Journal of Pragmatics 32 (2000) 1247-1252

goals and scope of discourse psychology are not exactly the same as those of the
various sister disciplines that investigate language structure and use. For example,
researchers in cognitive linguistics, text linguistics, computational linguistics, con-
versation analysis, and pragmatics typically have a strong background in linguistics,
but their theories are not strongly rooted in contemporary theories of cognition (see
Lehnert, 1997; Tomasello, 1998). In contrast, discourse psychologists have stronger
roots in cognitive science and cognitive psychology than in linguistics. In an ideal
world, there would be a tight harmony among psychological theories, linguistics
theories, discourse processing theories, corpora of naturalistic discourse excerpts,
and data from experiments on artificial texts (called 'textoids'; Graesser et al.,
1997). As it stands, however, the fields are currently more fractionated than in syn-
chrony.
In this book, Kintsch presents his 'construction-integration' (CI) model of text
comprehension, a model (theory) that was originally introduced by Kintsch a decade
ago (Kintsch, 1988). Kintsch reports the empirical evidence that tests the model and
explains how the CI model is more plausible psychologically than alternative com-
putational architectures. The CI model simulates the process of comprehending
printed text, although it could, in principle, be expanded to handle oral discourse. As
the name of the model suggestions, two major stages of processing are executed
when a sentence is comprehended and related to the prior discourse context: a con-
struction phase, followed by an integration phase.
During the construction phase, there is a quick activation of linguistic, semantic,
and world knowledge. This knowledge is triggered by the explicit words in the
incoming sentence (S) and the contents of working memory. Working memory is
limited in capacity. It includes a very small number of 'nodes' (i.e., approximately 5
or 6 words, ideas, or propositions) that are in the discourse focus or that are other-
wise highly activated just before sentence S is comprehended. These highly activated
nodes from sentence S and working memory (WM) trigger associated knowledge in
long-term memory. It takes approximately 400 milliseconds to activate the associ-
ated nodes from long-term memory (LTM). Some of this content from LTM refers
to previous nodes in the passage representation (which get reinstated) and some of
the content refers to generic content in the semantic and conceptual system. The con-
struction phase also activates syntactic rules for constructing propositions, rules for
interconnecting propositions (e.g., subordinate relations, direct relations), and rules
for constructing inferences (via transitivity or thematic structures). These rules and
nodes from LTM create a larger space of content elements, each of which varies in
strength of activation. Whereas sentence S and the focal content of WM may have
5-10 nodes, 50-100 nodes from LTM may end up being activated.
Multiple layers of nodes are activated during the construction phase. First, there is
the surface code, which captures the exact wording and syntax of the clauses. The
surface code hangs around only for a few seconds unless aspects of the surface code
have important repercussions on meaning. Second, there is the propositional
'textbase', which contains the explicit propositions in a stripped-down form that pre-
serves meaning, but not the surface code. Kintsch distinguishes atomic propositions,
which are predicate-argument constituents (e.g., the professor is young), from corn-
 Book review / Journal of Pragmatics 32 (2000) 1247-1252 1249

plex propositions, as illustrated below for the sentence 'the young professor ner-
vously opened his briefcase in the cafeteria yesterday'.

Category (action)
Predicate: open
Arguments (agent: professor, object: briefcase)
Argument modifier: possess (professor, briefcase)
Predicate modifier: nervously
Circumstance
Time: yesterday
Space: cafeteria

Third, there is a 'situation model' that refers to the content or microworld that the
text is about. The microworld in a story, for example, includes the setting and plot
that chronologically unfolds. Inferences are needed to flesh out the content of the sit-
uation model, although Kintsch sometimes hesitates to use the word 'inference'
when referring to the nonexplicit content that gets activated. One of the contempo-
rary debates in discourse psychology addresses the question of what inferences are
constructed in the situation model during comprehension (Graesser et al., 1994;
McKoon and Ratcliff, 1992; Singer, 1994; Zwaan and Radvansky, 1998).
Additional layers of nodes are needed to handle the broader context that situates
the act of comprehension. The reader may have particular goals for reading the text,
which would be implemented as a layer of nodes. The text genre would similarly be
captured by another layer of nodes. There are layers of nodes that Kintsch has not
yet implemented in his CI model, but they would presumably be needed in a com-
plete model of comprehension. Perhaps the status of these layers should be regarded
as challenges for future research. For example, there was no attempt to capture the
goals and attitudes of the writer, properties of the writer's anticipated audience, the
writer's intended meaning, specification of the common ground between writer and
reader, and multiple perspectives of discourse participants. Some models in dis-
course psychology have attempted to explain these pragmatic phenomena (Clark,
1996; Gibbs, in press). However, a rich theory of pragmatics and communication is
beyond the scope of what Kintsch has delivered in his current tests of the CI model,
It is tempting to say that pragmatics can be handled by adding additional layers of
nodes in the node space that gets built during the construction phase. Perhaps such
an augmentation of the CI model will work. If so, the researcher would need to add
the new layers of nodes in a principled fashion and to observe whether the CI model
ends up producing interesting patterns of data. However, there no doubt will be
skeptics among the researchers who study the pragmatics of discourse. To them,
adding the layer of pragmatic nodes would be regarded as an unilluminating ad hoc
exercise that misses the deep problems in pragmatics. Kintsch would undoubtedly
argue that the dispute should be settled by building and testing the augmented CI
model, not by arm-chair speculation.
During the 'integration' phase of the CI model, there is a complex dynamic pat-
tern of node activation that ends up converging on a coherent meaning representa-
1250 Book review / Journal of Pragmatics 32 (2000) 1247-1252

tion. This phase starts out with the large node space that was generated during the
construction phase, and a set of weights that connect all possible pairs of nodes in
this space. If the construction phase produces N nodes, there will be [N x (N - 1)]
weights in the weight space. Each weight specifies whether node X activates node Y
(positive weight value), as opposed to being inhibitory (negative weight) or unre-
lated (0 value). The values of these weights are generated theoretically in the best of
worlds, but sometimes are motivated by intuition or independent empirical data.
Once the matrix of weights is established, activation is spread throughout the net-
work according to the weights in the weight space, following the principles of con-
nectionist (neural network) architectures with parallel constraint satisfaction (Rumel-
hart and McClelland, 1986). The spreading activation continues over several
processing cycles until the connectionist network settles on a stable set of activation
values in the set of N nodes. An incoming sentence is easy to comprehend if few
cycles of processing are needed before the network settles. Difficult sentences
require several processing cycles, or oscillate between altemative activation patterns.
The activation strength of a particular node can be plotted as a function of time
during the integration phase. Some nodes get more active over time, such as those
that are connected to many other nodes with positive weights. Other nodes die out
because they are relatively isolated from other nodes or they have negative weights
with other nodes. In a similar fashion, it is possible to plot the activation level of a
particular node as a function of the evolving passage representation, sentence by sen-
tence. Memory for a node is predicted to increase as a function of the cumulative
activation across sentences in the text. Thus, the CI model furnishes complex pre-
dictions about the activation contours of individual nodes during the comprehension
of a particular sentence, and across successive sentences in the text. The complexity
and grain-size of this model outstrips most of the alternative models in discourse
psychology, although complex quantitative models are quite prevalent in the field
(Britton and Graesser, 1996).
One of the strengths of the CI model is the amount of empirical findings that it
can accommodate. Kintsch has tested the model on naturalistic stories and scientific
expository texts, as well as on mathematical word problems and experimenter-gen-
erated textoids that provide control over various content features. The experimental
tasks have tapped on-line processing (through reading times or decision latencies on
interspersed test words), off-line tests of memory representations (recall, sentence
recognition), statement verification, question answering, summarization, and deci-
sion making. He has compared readers with different amounts of expertise on the
topics covered in expository texts. The CI sometimes predicts complex interactions
between text, task, test, and reader expertise. For example, readers with low topic
knowledge perform better on coherent expository texts (compared to texts with
coherence gaps or organizational flaws) on virtually any test that taps the textbase
representation or the situation model representation. A different pattern emerges for
readers with high topic knowledge. Their performance on the tests on the textbase
are slightly better for coherent texts than texts with coherence gaps, but it is the low
coherence texts that reign supreme in tests that tap the situation model. Coherence
gaps and poor text organization force the high knowledge reader to construct a richer
 Book review / Journal of Pragmatics 32 (2000) 1247-1252 1251

situation model, but low knowledge readers have difficulty doing this because they
have a skimpy base of world knowledge.
Kintsch argues that there are a number models in the sister sciences that are not
psychologically plausible. What are some of these implausible models? The human
mind does not have a front-end syntactic, rule-based, parser that is a separate mod-
ule, independent of the lexicon and world knowledge. The human mind does not
have a symbolic reasoning engine that performs predicate calculus. The human mind
does not generate multiple worlds when a sentence is comprehended and does not
forecast alternative future plots in the midst of comprehending a story. The human
mind does not have brittle rules and categories. According to Kintsch, the human
mind settles on a single meaning representation when a sentence is comprehended in
a text, following a connectionist model with multiple interacting layers of nodes and
soft constraint satisfaction.
It will be interesting to see whether Kintsch's CI model will end up providing
incisive solutions to traditional problems in mainstream pragmatics. For example,
what would it take to handle sarcasm, politeness, indirect requests, intended vague-
ness, multiple voices, and embedded discourse spaces? What additional layers of
nodes will be needed and how will the nodes be connected in the weight space? Will
the resulting patterns of node activation be insightful and provide explanatory fits to
the data? Research on these questions is one way to strengthen the communication
between discourse psychology and contemporary pragmatics.

References
Britton, Bruce K. and Arthur C. Graesser, eds., 1996. Models of understanding text. Mahwah, NJ: Erl-
baum.
Clark, Herbert H., 1996. Using language. Cambridge: Cambridge University Press.
Gibbs, Raymond Jr., in press. Intentions in the experience of meaning. New York: Cambridge Univer-
sity Press.
Graesser, Arthur C., Keith K. Millis and Rolf A. Zwaan, 1997. Discourse comprehension. Annual
Review of Psychology 48: 163-189.
Graesser, Arthur C., Murray Singer and Tom Trabasso, 1994. Constructing inferences during narrative
text comprehension. Psychological Review 101:371-395.
Kintsch, Walter, 1988. The role of knowledge in discourse comprehension: A construction-integration
model. Psychological Review 95:163-182.
Lehnert, Wendy, 1997. Information extraction: What have we learned? Discourse Processes 23:
441-470.
McKoon, Gail and Roger Ratcliff, 1992. Inference during reading. Psychological Review 99: 440--466.
Rumelhart, David E. and James L. McClelland, 1986. Parallel distributed processing: Explorations in the
microstructure of cognition, Volume 1. Cambridge, MA: MIT Press.
Singer, Murray, 1994. Discourse inference processes. In: M.A. Gemsbacher, ed., Handbook of psy-
cholinguistics, 479-515. San Diego, CA: Academic Press.
Tomasello, M., ed., 1998. The new psychology of language. Mahwah, NJ: Erlbaum.
Zwaan, Rolf A. and Gabriel A. Radvansky, 1998. Situation models in language comprehension and
memory. Psychological Bulletin 123 : 162-185.
1252 Book review / Journal of Pragmatics 32 (2000) 1247-1252

Dr. Graesser is presently a full professor in the Department of Psychology and an adjunct professor in
Mathematical Sciences at The University of Memphis. He is currently a co-director of the Institute for
Intelligent Systems and director of the Center for Applied Psychological Research. In 1977 Dr. Graesser
received his Ph.D. in psychology from the University of California at San Diego. He was a visiting
researcher at Yale University in 1983, Stanford University in 1984, and Carnegie Mellon University in
1991. Dr. Graesser's primary research interests are in cognitive science and discourse processing. More
specific interests include knowledge representation, question asking and answering, tutoring, text com-
prehension, inference generation, conversation, reading, education, memory, expert systems, artificial
intelligence, and human-computer interaction. He is currently editor of the journal Discourse Processes.
In addition to publishing approximately 190 articles in journals and books, he has written two books and
has edited six books.

Shannon Whitten is a graduate student in the cognitive science lab at the University of Memphis. Her
principle research interests are in the area of discourse comprehension, specifically thematic inferences
and perception of time in situation models. She received her B.A. at Florida State University.