Quantification of Stylistic Traits:

A Statistical Approach

Mappillairaju Bagavandas, G. Manimannan

Department of Statistics – Madras Christian College – Chennai-600 059 – India
mbdas49@hotmail.com, manimannang@yahoo.co.in

Abstract
It is often recognized that authors have writing styles and it is possible to find a simple statistical mo del, which
explains reasonably what makes an author unique. This paper makes an attempt to identify the distinct stylistic
features of three Tamil Scholars of the same period and also tries to quantify the writing styles of these authors
using eighteen stylistic features. These stylistic features of this study are eleven morphological variables, four
habitual words and three function words. ANOVA technique, two sample t-statistic and Factor analysis are used
for measuring such stylistics traits and also identifies those traits, which are most densely packed.

Keywords: author style statistics, ANOVA, two-sample t-statistic, factor analysis.

1. Introduction
It has been recognized that an author has a unique writing style which is expressed in the form
of subconscious stylistic features. Style of an author can be quantified by counting his\her
choice of words for expressing his\ her ideas under the assumption that the writer favouring a
stock of words for the expression of ideas is regarded, to some extent, subject to chance
(Holmes and Forsyth, 1995). Hence given a certain personality and thus a certain style, as its
expression, the characteristic properties of style can be described in terms of statistical law
(Herdan, 1964). Bailey (1979) says that the stylistic features of a matured writer will be sali-
ent, structural, frequent and easily quantifiable. Thus style reflects personality of a writer and
this unconscious process is consistent in the case of matured writers (Holmes, 1985).
Statistical stylistic study not only compliments the traditional scholarship of literary experts
but also provides an alternative method for investigating the works of doubtful provenance
(Holmes, 1998). These studies provide authentic results if they work within the same genre
and also work within as close a time period as possible. Stylistic markers which occur most
frequently in a given passage are also identified by these methods (Mealand, 1997). These
stylistic studies inhabit two types of problems, the first being the selection of suitable set of
stylistic variables and the second being the selection of appropriate techniques. There is no
general agreement on the stylistic variable that should be used in stylistic studies. In general,
when choosing the stylistic variables, one must use something that has large variation across
authors and relatively little variation among an author's own work. Initially, lexical variables
have predominated in the stylometry studies, yet this decade has seen the application of
syntactic and semantic variables (Holmes, 1998).
Mathematician like Fucks (1952) may be considered pioneers in laying a foundation for more
vigorous and objective stylistic analysis through his attempts to quantify stylistic features.
Mosteller and Wallace's study (1964) is considered as the first authentic stylometric study

JADT 2004 : 7es Journées internationales d’Analyse statistique des Données Textuelles
72 MAPPILLAIRAJU BAGAVANDAS, G. MANIMANNAN

soundly based on modern statistical procedure using computer as its major research tool. John
Burrows (1987) through his series of seminal papers introduced stylometry studies as a viable
tool for authorship attribution problem. The availability of modern computing facility has
provided a unique opportunity for many stylometricians to introduce many multivariate meth
ods like factor analysis, cluster analysis and correspondence analysis for conducting experi-
ments with high dimensional data and also to widen the frontiers of stylometry (Peng, 2001).
Factor analysis is considered as an ideal method for determining the relationship between sty-
listic features and stable personality traits (Sommers, 1966; Herdan, 1964). This analysis helps
to find out whether different writers really represent different distinct forms of behaviours or
whet her they draw from a limited stock of vocabulary (Miles and Selvin, 1966). This
multivariate technique is also used for measuring the extent to which groups of words have
similar patterns of high or low use of various writers.
Herdan (1941) was the first to use the factor analysis for analysing the relation between six
authors and for identifying the one who uses the most difficult words. This analysis was also
applied to establish the common ancestors of a number of proto Indo-European languages
(Johnson and Kotz, 1967). Roger Peng and Nicolas Hengartner (2002) have used factor analy-
sis to examine each individual author's function-word counts and also to filter out words
which account for very little of the variation between authors in the group. This analysis was
used by David Mealand (1997) to establish that samples of different genres from the Gospel
of Mark vary in style and also to identify the stylistic markers which are most heavily used in
these passages.

2. Data and methods

The present study deals with the literary works of three contemporary Tamil scholars,namely,
MahakaviBarathi (MB), V.Kalyanasundaram (VK) and Subramaniya Iyer (SI). In the Pre–In-
dependence period, these three scholars have written number of articles on India's Freedom
Movement in the magazine called India. Initially, all the three scholars have written articles by
attributing their names. The oppressive attitude of the then British Regime made all the three
writers to write articles on the same topic anonymously in the same magazine. All the attrib-
uted and unattributed articles written on India’s Freedom Movement in that magazine were
complied and brought out as a book entitled Bharathi Dharisanam in the year 1975. For this
quantitative stylistic study, all attributed articles of these three scholars written on India’s
Freedom Movement in the year 1906 are considered. Our study is based on nineteen articles of
Bharati, six of Kalyanasundram and seven of Subramaniya Iyer.
In stylometry, there are important decisions to be made about the features to be selected and
the methods to be used (Mealand, 1997). Eighteen stylistic features are consider ed for this
study. They include eleven morphological variables, four habitual words and three function
words. The exact lists of variables of this study with their abbreviations are given in Table 1.
For a comparative analysis the frequency counts of the stylistic features must be normalized to
the text length in an article. In this study since each sentence is considered as a sample, to
normalize the stylistic features, the raw frequency counts of each stylistic feature is divided by
the number of words in each sentence and then multiplied by hundred to express it in percent-
age. Eighteen stylistic features are identified from each sentence. These features include parts
of speech, habitual words and function words. Both voices and tenses are expressed in fre-
quencies but not in percentages. If we have n sentences and if we identify p stylistic features

JADT 2004 : 7es Journées internationales d’Analyse statistique des Données Textuelles
 QUANTIFICATION OF STYLISTIC TRAITS: A STATISTICAL APPROACH 73

from each sentence, then we have a data matrix of size n x p. Thus, each article was converted
as a data matrix and these data matrices form the basis for this quantitative study.
A chi-square analysis of the nineteen articles of Bhararthi establishes that these articles do not
differ from one another in terms of the frequency distribution of occurrence of these stylistic
features. Similar results were obtained in the case of other two scholars (Manimannan and
Bagavandas, 2001). Hence all the nineteen articles of Bharathi are considered as one article
for this study. So also, the six articles of Kalyanasundram and seven articles of Subramaniya
Iyer. In this study, each sentence is considered as a sample. Hence the nineteen articles of
Bharathi consist of three hundred and fifty three sentences, six articles of Kalyanasundram
consist three hundred and eighty two sentences and seven articles of Subramaniya Iyer consist
of three hundred and fifteen sentences. As there are three authors, there are three data matrices
and their sizes are (353x18), (382x18) and (315x18) respectively. Hence the aim is to
compare the data matrices of the linguistic features of the three scholars. Average values, two-
sample t-statistic values and Euclidean distance values are given in Table1.

3. Analysis
This analysis section consists of two parts. Part one identifies the special stylistics features of
each author and Part two quantifies the writing style of each author.
3.1. Identification of Special Stylistic Features
This univariate analysis compares the average values of the stylistic features of the three
scholars. This comparative study is made in two stages. In the first stage, the hypothesis of
equality of the means of a particular feature of three authors is tested using ANOVA (one-
way) technique. The acceptance of this hypothesis indicates that particular stylistic feature has
no discriminatory power. However, if this hypothesis is rejected, then mean difference of a
feature between any two authors is tested using the conventional two-sample t-statistic.
This two-stage comparative analysis indicates that the stylistic features like two-letter word,
three-letter word and pronoun do not discriminate these three scholars from one another. That
is, these three scholars had the habit of using the same number of these features in writing a
sentence. This result indicates that all these three scholars had used, on an average, one pro-
noun, one two-letter word and two three-letter words in a sentence of ten words. Also the
smaller percentages of occurrence of features like intensifier, infinity and adverb indicates that
these three authors had used these three features very rarely.
The percentages of occurrences of stylistic features like noun, post-position, clitic, case mak-
ers and conjunctions differentiate these three authors statistically from one another. This result
indicates that Bharathi is identified as the least user of these features whereas Kalyanasundram
is identified as the maximum user of the same stylistic features. Subramaniya Iyer is not iden-
tified with any distinct stylistic features because the percentages of the occurrences of stylistic
features of this author indicate that the writing style of this author shares equally the special
features of the other two authors. The Euclidean distance values confirm this result in Table 1.
This analysis shows that in the sentence of ten words, Bharathi had used, on an average, one
postposition, one clitic but three nouns and four case markers and two conjunctions. But on
the other hand, in the sentence of the same length, on an average, Kalyanasundram had used
five nouns, four conjunctions, three postpositions, three clitics but six case markers. Also it
can be seen that the third author, Subramania Iyar had used, on an average, four nouns, three
post positions, three clitics, six case markers and four conjunctions.

JADT 2004 : 7es Journées internationales d’Analyse statistique des Données Textuelles
74 MAPPILLAIRAJU BAGAVANDAS, G. MANIMANNAN

3.2. Stylo – Statistical Analysis

Factor analysis is a variable-oriented multivariate technique. This analysis describes the inter-
relationship among many variables in terms of a few underlying, but observable, random
qualities called factors (Lawley and Maxwell, 1971). Factor analysis can be considered as an
extension of principal component analysis and is used for data reduction and interpretation.
This analysis is also used for grouping of variables in such a way that the variables are highly
correlated with in groups but have relatively insignificant correlation with variables of differ-
ent groups. Correlation matrix of the eighteen stylistic features is calculated for each data ma-
trix. The initial statistics are given in Table 2 and groups of stylistic features are given in Ta-
ble 3.
3.2.1. The case of Bharathi
All the eighteen features are highly loaded in the first seven factors, which covers nearly 54 %
of the total variation present in this data set. In other words these eighteen features are
grouped into seven clusters on the basis of the inter-relationship among themselves. The fea-
tures like words starting with vowel, verb, two-letter, three-letter and four-letter words are
highly loaded in the first factor and hence they form as a cluster. This result shows that the
writer Bharathi had preferred verbs and words starting with vowels either as two-letter or
three-letter or four-letter words. Since four out of these five features are habitual words, this
factor is named as habitual-word factor.
Factor two is highly correlated with features like clitics and case makers. These correlated
relationships establish that this writer had the habit of using clitics and case makers in the
ratio of 1: 4 in a sentence of ten words. As these two features are function words, this factor is
known as function-word factor. Third factor is a contrast between features like noun and pro-
noun and also they occur in the ratio 4:1 and whenever the occurrence of noun increases the
occurrence of pronoun decreases in a sentence. These two features are morphological vari-
ables and hence this factor is known as morphological factor. Statistical features like tenses
and numeral are accommodated in the fourth factor-the tense factor. This factor indicates that
this writer had used to write sentences mostly in past tense with a very few numerals.
Since features like voice and postposition are accommodated in the fifth factor, this factor
may be named as voice factor. This factor is a contrast between voice and postposition. This
indicates that the writer had favoured to write sentences in the past tense with less number of
postpositions. The sixth factor is a syllable factor and it’s established that the length of ten
words sentence on an average fifteen syllables.
The Seventh factor is contrast between two groups of features. Infinity and adverb are grouped
together and intensifiers and conjunction are grouped together. The occurrence of these stylis-
tic features like intensifier, infinity and adverb are rare phenomena. But Bharathi had used at
least two conjunctions in a sentence of ten words and hence this factor may be called
conjunction factor.
Summarizing, the writer Bharathi had used passive voice sentences in past tense to narrate
India’s Freedom Movement. In sentence of ten words, he had used, on the average, one clitic,
one pronoun, two verbs, three words starting with vowels, four case makers and four nouns.
The verbs and words starting with vowels are either two-letter or three-letter or four-letter
words. The incre ase in the occurrence of nouns reduces the occurrence of pronouns.

JADT 2004 : 7es Journées internationales d’Analyse statistique des Données Textuelles
 QUANTIFICATION OF STYLISTIC TRAITS: A STATISTICAL APPROACH 75

3.2.2. The case of Kalyanasundram

The first four factors, which cover nearly 36 % of the total variation present in the data set,
had grouped all eighteen features into four clusters. In the first factor features like case maker
and clitic are highly accommodated in the ratio 3:1 and hence this factor is known as function-
word factor. The features like verb and noun are grouped in the second factor in the ratio 1: 2.
This factor is a contrast between verb and noun, which indicates that whenever the occurrence
of nouns increases, the occurrence of verbs decreases. This is a morphological factor.
The third factor is a habitual-word factor as it accommodates all the three habitual words with
pronoun. This factor is a contrast between four-letter word and the set of two-letter and three-
letter words and pronouns. More the occurrence of pronouns, less it will be four-letter words.
Fourth factor is a contrast factor. Adverb, syllable, conjunction and infinity are grouped in one
set and voice, tense, word starting with vowel, intensifier, and clitics are grouped in another
set. This result shows that this author had written active voice sentence in past tense. This
author has provided four conjunctions, three words starting with vowels and three clitics. The
occurrence of more conjunctions in a sentence reduces the occurrence of clitics and words
starting with vowels.
Finally, the author Kalyanasundram used active voice sentences in past tense to describe
India’s Freedom Movement. In these sentences of ten words, on the average, three postposi-
tions, three clitics, three words starting with vowels, four conjunctions, four nouns and six
case markers are accommodated. The occurrence of more verbs red uces the occurrence of
nouns; also the occurrence of more conjunctions reduces the occurrences of clitics and the
words starting with vowels.
3.2.3. The case of Subramaniya Iyar
All the eighteen features are accommodated in the first seven factors, which covers nearly
56 % of variation present in the given data set. In the first factor, verb, word starting with
vowel, syllable and four-letter word are accommodated. This indicates that verb and word
starting with vowel will be four-letter words with two or three syllables. This is a morphologi-
cal factor. Case marker and clitic are highly loaded in the second factor and they occur in the
ratio 2:1 in a sentence. This is a function- word factor.
Third factor, a contrast factor, provides high loading for nouns and pronouns in the ratio 6:1.
The occurrence of more nouns reduces the occurrence of pronouns. This is a noun-family
factor. Fourth factor accommodates voice and conjunction. This author used to write passive
voice sentences with at least three conjunctions. In the fourth factor adverb and two-letter
word are accommodated and this indicates that the adverbs of this author are identified as
two-letter words.
Fifth factor is a tense factor. This writer used to write sentence in present tense. The last factor
contrasts between two sets of features. In one set postposition, intensifier and three-letter word
are accommodated and in the other set numeral and infinity are accommodated. This result
indicates that there will be three post-positions and two three-letter words in a sentence of ten
words.
Summarizing, the scholar Subramaniya Iyar made use of passive voice sentences in present
tense. There will be six case markers, four nouns, three conjunctions and three postpositions
and one pronoun in a sentence. The verb and word starting with vowel will be four-letter
words with two or three syllables. The occurrences of more nouns reduce the occurrence of
pronouns.

JADT 2004 : 7es Journées internationales d’Analyse statistique des Données Textuelles
76 MAPPILLAIRAJU BAGAVANDAS, G. MANIMANNAN

4. Conclusions
This study provides opportunities to introduce statistical techniques for identifying the special
stylistic features and also for quantifying the writing styles of three Tamil scholars, namely,
Mahakavi Bharathi, V. Kalyanasundram and Subramania Iyer using eighteen stylistic features.
Articles written on India’s Freedom Movement by these scholars are considered for this study.
Bharathi had written sentences in past tense with the least function words. V. Kalyanasundram
is identified as a writer who has used maximum number of function words in active voice
sentences with past tense. The third writer, Subramaniya Iyar has written sentences in passive
voice but in present tense and is not identified with any distinct stylistic features.

References
Bailey R.W. (1979). The Future of Computational Stylistic. Association for Literary and Linguistic
Computing Bulletin.
Burrow J. (1987). Word-Patterns and Story-shapes: The Statistical Analysis of Narrative Style. Liter-
ary and Linguistic Computing, vol. (2/2): 61-70.
Fucks W. (1952). On the Mathematical Analysis of Style. Biometrica, vol. (39): 122-129.
Herdan G. (1941). The Advanced Theory of Language as Choice and Chance. The Hegue.
Herdan G. (1964). On Communication between Linguistics. Linguistics, vol. (9): 71-76.
Holmes D.I. (1985). The Analysis of Literary Style: A Review. Journal of the Royal Statistical Soci-
ety, Series A (155): 91-120.
Holmes D.I. and Forsyth R.S. (1995). The Federalist Revisited: New Directions in Authorship Attri-
bution. Literary and Linguistic Computing, vol. (10): 11-27.
Holmes D.I. (1998). The Evolution of Stylometry in Humanities Scholarship. Literary and Linguistic
Computing, vol. (13): 111-117.
Johnson N.L. and Kotz. S. (1967). Discrete Distributions. Houghton Miffin Company Boston.
Lawley P.A. and Maxwell. A.E. (1971). Factor Analysis as a Statistical Method (2nd Ed.). American
Elsevier Publishing and Co. New York.
Manimannan G. and Bagavandas M. (2001). The Authorship Attribution: the case of Bharathiyar. In
Paper Presented at National Conference on Mathematical and Applied Statistics, Nagpur Univer-
sity.
Mealand D. (1997). Measuring Genre Differences in Mark with Correspondence Analysis. Literary
and Linguistic Computing, vol. (12/4).
Miles J. and Selvin H.C. (1966). A Factor Analysis of the Vocabulary of Property in the Seventeenth
Century. In Leed J. (Ed.), The Computer and Literary Style. State University Press.
Mosteller F. and Wallace D.L. (1964). Applied Bayesian and Classical Inference, The Case of the
Federalist Papers. Addition-Wesley, Reeding.
Peng R. and Hengartner N. (2001). Quantitative Analysis of Literary Style. University of California,
CA90095.
Peng R. and Hengartner N. (2001). Statistical Aspects of Literary Style. Yale University, CC-99.

JADT 2004 : 7es Journées internationales d’Analyse statistique des Données Textuelles
 QUANTIFICATION OF STYLISTIC TRAITS: A STATISTICAL APPROACH 77

Two samples t-statistic

Abbre Mean values Euclidean values
Stylistic features values
viations
MB VK SI MB-VK MB-SI VK-SI MB-VK MB-SI VK-SI

Noun P_Noun 34.26 45.47 41.8 09.38 06.21 02.91 125.49 56.73 13.47

Intensifier P_Int 00.05 06.07 06.1 11.63 10.31 00.09* 31.07 31.79 0.00

Infinitive P_Inf 00.58 01.33 03.6 02.69 06.14 04.26 0.57 9.18 5.19

Pronoun P_Pro 07.72 07.73 06.6 00.01* 01.55* 01.64* 0.00 1.19 1.20

Tense Tense 01.71 01.77 01.4 01.32* 05.84 08.42 0.00 0.08 0.11

Numeral P_Nume 03.99 05.27 05.3 02.26 02.20 00.13* 1.62 1.83 0.01

Two-Letter Word P_Two 10.61 09.79 09.5 01.04* 01.31* 00.31* 0.68 1.13 0.06

Three-Letter Word P_Thre 19.24 20.18 18.3 00.86* 00.69* 01.52* 0.88 0.78 3.32

Four-Letter word P_Four 20.46 25.66 25.9 04.20 03.86 00.19* 2.08 30.0 0.07

Word starting with vowels P_Vowe 27.74 33.36 28.8 04.09 00.72* 03.35 31.61 1.27 20.22

Verb P_Verb 23.43 21.40 23.6 02.52 00.22* 02.19 4.13 0.06 5.18

Voice Voices 02.25 01.55 02.1 10.88 01.69* 08.69 0.49 0.01 0.34

Syllable P_Sylla 151.10 119.70 163.0 02.82 00.88* 03.40 989.31 161.21 1949.20

Post position P_Post 13.43 35.26 31.3 15.00 11.68 02.06 476.38 322.53 14.95

Clitics P_Clitic 14.14 34.25 33.4 16.31 14.93 00.52* 404.28 371.69 0.68

Case marker P_Case 38.65 69.95 63.0 15.15 12.45 02.75 980.00 596.68 47.30

Adverb P_Adverb 04.39 02.26 02.8 04.22 02.78 01.29* 4.54 2.37 0.35

Conjunction P_Conjun 22.65 42.15 35.5 11.11 07.42 03.48 380.24 165.74 43.90

Total 3458.37 1754.2 2105.60

SQRT 58.81 41.88 45.89

* not significance at 5% level

Table 1. Mean value, Two-samples t-statistic values and Euclidean distance values

JADT 2004 : 7es Journées internationales d’Analyse statistique des Données Textuelles
78 MAPPILLAIRAJU BAGAVANDAS, G. MANIMANNAN

MB VK SI
Factors Eigen Percentage Cumulative Eigen Percentage Cumulative Eigen Percentage Cumulative
values of variance percentage values of variance percentage values of variance percentage
1 2.223 12.3 12.3 2.018 11.2 11.2 2.492 13.8 13.8
2 1.560 8.7 21.0 1.645 9.1 20.3 1.590 8.8 22.7
3 1.420 7.9 28.9 1.463 8.1 28.5 1.483 8.2 30.9
4 1.270 7.1 36.0 1.329 7.4 35.9 1.217 6.8 37.7
5 1.167 6.5 42.4 1.150 6.4 42.3 1.135 6.3 44.0
6 1.103 6.1 48.6 1.068 5.9 48.2 1.094 6.1 50.1
7 1.035 5.7 54.3 1.048 5.8 54.0 1.008 5.6 55.7
8 1.013 5.6 59.9 1.004 5.6 59.6 0.966 5.4 61.0
9 0.961 5.3 65.3 0.943 5.2 64.8 0.903 5.0 66.0
10 0.945 5.3 70.5 0.916 5.1 69.9 0.876 4.9 70.9
11 0.856 4.8 75.3 0.843 4.7 74.6 0.829 4.6 75.5
12 0.837 4.6 79.9 0.784 4.4 78.9 0.772 4.3 79.8
13 0.758 4.2 84.1 0.745 4.1 83.1 0.720 4.0 83.8
14 0.702 3.9 88.0 0.738 4.1 87.2 0.673 3.7 87.6
15 0.646 3.6 91.6 0.685 3.8 91.0 0.658 3.7 91.2
16 0.546 3.0 94.7 0.580 3.2 94.2 0.597 3.3 94.5
17 0.450 2.8 97.4 0.540 3.0 97.2 0.538 3.0 97.5
18 0.460 2.6 100.0 0.502 2.8 100.0 0.447 2.5 100.0

Table 2. Factor analysis-Initial statistics

Factors MB VK SI
FACTOR 1 (.75105) P-VOWE (.65782) P_CASE (.71539) P_VERB
(.67532) P-VERB (.56787) P_POST (.54890) P_FOUR
(.55204) P-TWO (.64105) P_SYLLA
(.58042) P-THRE (.34474) P_VOWE
(.66074) P-FOUR
FACTOR 2 (.83945) P_CLITIC (-.53092) P_NOUN (.77953) P_CASE
(.85563) P_CASE (.73761) P_NUME (.76640) P_CLITIC
(-.68164) P_VERB
FACTOR 3 (.70597) P_NOUN (.75332) P_FOUR (-.66287) P_PRO
(.82167) P_PRO (-.67378) P_THRE (.62711) P_NOUN
(.59352) P_TWO
(.75897) P_PRO
FACTOR 4 (-.66491) TENSE (.79795) P_ADVERB (-.71236) VOICE
(.69586) P_NUME (-.39408) TENSES (.59945) P_CONJON
(.83215) P_SYLLA
(.66924) P_CONJON
(.71794) VOICE
(.44038) P_VOWE
(.73773) P_INF
(.46984) P_INT
(.56534) P_CLITIC
FACTOR 5 (.73510) VOICE (.79019) P_TWO
(.62676) P_POST (.56156) P_ADVERB
FACTOR 6 (.66129) P_SYLLA (-.76712) TENSE
FACTOR 7 (.18315) P_INT (.10776) P_INF
(.66689) P_INF (.61875) P_NUME
(.57019) P_ADVERB (.39798) P_POST
(-.52044) P_CONJON (.76640) P_CLITIC
(.72258) P_THRE

Factor scores are given in the brackets

Table 3. Grouping of stylistic features according to factor scores

JADT 2004 : 7es Journées internationales d’Analyse statistique des Données Textuelles