Some forty years ago, Mosteller and Wallace suggested in their influential work on the Federalist
Papers that a small number of the most frequent words in a language ('function words')
could usefully serve as indicators of authorial style. The decades since have seen this work taken
up in many ways including both the use of new analysis techniques (discriminant analysis, PCA,
neural networks, and more), as well as the search for more sophisticated features by which to
capture stylistic properties of texts. Interestingly, while use of more sophisticated models and
algorithms has often led to more reliable and generally applicable results, it has proven quite
difficult to improve on the general usefulness of function words for stylistic attribution. Indeed,
John F. Burrows, in his seminal work on Jane Austen, has demonstrated that function
words can be quite effectively used for attributing text passages to different authors, novels, or
individual characters.
The intuition behind the utility of function words for stylistic attribution is as follows.
Due to their high frequency in the language and highly grammaticalized roles, function words are
very unlikely to be subject to conscious control by the author. At the same time, the frequencies
of different function words vary greatly across different authors and genres of text - hence the
expectation that modeling the interdependence of different function word frequencies with style
will result in effective attribution. However, the highly reductionistic nature of such features
seems unsatisfying, as they rarely give good insight into underlying stylistic issues, hence the
various efforts at developing more complex textual features while respecting constraints on
computational feasibility.
One especially promising line of work in this regard has been the examination of frequent
word sequences and collocations for stylistic attribution, particularly Hoover's recent (2004)
systematic work on clustering analysis of several text collections using frequent word
collocations. A "word collocation" is defined as a certain pair of words occurring within a given
threshold distance of each other (such as "is" and "certain" appearing within 5 words of each
other in this sentence). Given such a threshold, the most frequent such collocations are
determined over the entire corpus, and their frequencies in each text constitute its features for
analysis. Hoover's analyses show the superiority, for his data set, of using frequent word
collocations (for certain window sizes) over using frequent words or pairs of adjacent words.
We contend, however, that by using such a small data set (twenty samples of 10,000
words each, in one case), the discriminating power of a model based on function words will be
much reduced, and so the comparison may not be fair. As has been shown for other
computational linguistic tasks (see, e.g., Banko & Brill), even simple language modeling
techniques can greatly improve in effectiveness when larger quantities of data are applied. We
have therefore explored the relative effectiveness of frequent words compared to frequent pairs
and collocations, for attribution of both author identity and national origin, increasing the number
of text passages considered over earlier work.
We performed classification experiments on the twenty novels considered by Hoover,
treating each separate chapter of each book as a separate text (rather than using just the first
10,000 words of each novel as a single text). Table 1 gives the full list with numbers of chapters
and average number of words per chapter. We used a standard state-of-the-art machine learning
technique to derive linear discrimination models between pairs of authors. This procedure gave
results that clearly show a superiority of function words over collocations as stylistic features.
Qualitatively similar results were obtained for the two-class problem of attributing the national
origin (American or British) of a text's author. We conclude from this that larger and more
detailed studies need to be done to effectively validate the use of a given feature type for
authorship attribution.
| Author |
Book |
# Chapters |
Avg. Words |
| Cather |
My Antonia
|
45 |
1826 |
|
Song of the Lark
|
60 |
2581 |
|
The Professor's House
|
28 |
2172 |
| Conrad |
Lord Jim
|
45 |
2913 |
|
The Nigger of the Narcissus
|
5 |
10592 |
| Hardy |
Jude the Obscure
|
53 |
2765 |
|
The Mayor of Casterbridge
|
45 |
2615 |
|
Tess of the d'Urbervilles
|
58 |
2605 |
| James |
The Europeans
|
12 |
5003 |
|
The Ambassadors
|
36 |
4584 |
| Kipling |
The Jungle Book
|
13 |
3980 |
|
Kim
|
15 |
7167 |
| Lewis |
Babbit
|
34 |
3693 |
|
Main Street
|
34 |
4994 |
|
Our Mr. Wrenn
|
19 |
4126 |
| London |
The Call of The Wild
|
7 |
4589 |
|
The Sea Wolf
|
39 |
2739 |
|
White Fang
|
25 |
2917 |
| Wells |
The Invisible Man
|
28 |
1756 |
|
The War Of The Worlds
|
27 |
2241 |
Table 1. Corpus composition.
Given each particular feature set (frequent words, pairs, or collocations), the method was to
represent each document as a numerical vector, each of whose elements is the frequency of a
particular feature of the text. We then applied the SMO learning algorithm (Platt) with
default parameters, which gives a model linearly weighting the various text features. SMO is a
support vector machine (SVM) algorithm; SVMs have been applied successfully to a wide variety
of text categorization problems (Joachims).
Generalization accuracy was measured using 20-fold cross-validation, in which the 633
chapters were divided into 20 subsets of nearly equal size (3 or 4 texts per subset). Training was
performed 20 times, each time leaving out one of the subsets, and then using the omitted subset
for testing. The overall classification error rate was estimated as the average error rate over all 20
runs. This method gives a reasonable estimate of the expected error rate of the learning method
for each given feature set and target task (Goutte).
Results of measuring generalization accuracy for different feature sets are summarized in Tables
2 and 3, which clearly show that using the most frequent words in the corpus as features for
stylistic text classification gives the highest overall discrimination for both author and nationality
attribution tasks.
| Feature Set |
Author |
Nationality |
| Freq. Words |
99.00%
|
93.50%
|
| Freq. Pairs |
91.60% |
91.30% |
| Freq. Coll. (k=5) |
88.94% |
90.20% |
| Freq. Coll. (k=10) |
84.00% |
87.20% |
Table 2. 20-fold cross-validation results for 200 most frequent words, pairs, and collocations
(window size k = 5 or 10).
| Feature Set |
Author |
Nationality |
| Freq. Words |
93.20% |
93.50%
|
| Freq. Pairs |
90.00% |
88.60% |
| Freq. Coll. (k=5) |
91.50% |
92.10% |
| Freq. Coll. (k=10) |
94.00% |
92.10% |
Table 3. 20-fold cross-validation results for 500 most frequent words, pairs, and collocations
(window size 5 or 10).
Our study here reinforces many others over the years in showing the surprising resilience of
frequently-occurring words as indicators of the stylistic character of a text. Our results show
frequent words enabling more accurate text attribution than features such as word pairs or
collocations, surprisingly contradicting recent results as well as the intuition that pairs or
collocations should be more informative. The success of this study at showing the power of
frequent words we mainly attribute to the use of more data, in the form of entire novels, broken
down by chapters. The more fine-grained breakdown of text samples for each author enables
more accurate determination of a good decision surface for the problem, thus better utilizing the
power of all features in the feature set. Furthermore, using more training texts than features
seriously reduces the likelihood of overfitting the model to the training data, improving the
reliability of results.
It is indeed possible that collocations may be better than function words for different
stylistic classification tasks; however such a claim remains to be proven. A more general
interpretation of our results is that since a set of frequent collocations of a given size will contain
fewer different words than a set of frequent words of the same size, it may possess less
discriminatory power. At the same time, though, such a feature set will be less subject to
overfitting, and so may appear better when very small sets of texts are studied (as in previous
studies). Our results thus lead us to believe that most of the discriminating power of collocations
is due to the frequent words they contain (and not the collocations themselves), thus frequent
words outperformed collocations, given sufficient data.
Function words still prove surprisingly useful as features for stylistic text attribution, even after
many decades of research on features and algorithms for stylometric analysis. We believe that
significant progress is likely to come from fundamental advances in computational linguistics
which allow automated extraction of more linguistically motivated features, such as recent work
on extracting rhetorical relations in a text (Marcu).
More generally, our results argue for the importance of using larger data sets for
evaluating the relative utility of different attribution feature sets or techniques. As in our case of
comparing frequent words with frequent collocations, changing the scale of the data set may
affect the relative power of different techniques, thus leading to different conclusions. We
suggest that the authorship attribution community should now work towards developing a large
suite of corpora and testbed tasks, to allow more rigorous and standardized comparisons of
alternative approaches.