Comparative evaluation of term selection functions for authorship attribution

Abstract

Different computational models have been proposed to automatically determine the most probable author of a disputed text (authorship attribution). These models can be viewed as special approaches in the text categorization domain. In this perspective, in a first step we need to determine the most effective features (words, punctuation symbols, part-of-speech, bigram of words, etc.) to discriminate between different authors. To achieve this, we can consider different independent feature-scoring selection functions (information gain, gain ratio, pointwise mutual information, odds ratio, chi-square, bi-normal separation, GSS, Darmstadt Indexing Approach (DIA), and correlation coefficient). Other term selection strategies have also been suggested in specific authorship attribution studies. To compare these two families of selection procedures, we have extracted articles from two newspapers and belonging to two categories (sports and politics). To enlarge the basis of our evaluations, we have chosen one newspaper written in the English language ('Glasgow Herald') and a second one in Italian ('La Stampa'). The resulting collections contain from 987 to 2,036 articles written by four to ten columnists. Using the Kullback-Leibler divergence, the chisquare measure and the Delta rule as attribution schemes, this study found that some simple selection strategies (based on occurrence frequency or document frequency) may produce similar, and sometimes better, results compared with more complex ones.

A simple and efficient algorithm for authorship verification

“…Those terms are selected for the disputed text. For determining the value of k , previous studies have shown that a value between 200 and 300 tends to provide the best performance (Burrows, ; Savoy, ). This reduced number represents a huge difference compared to the 100,000 features used by Koppel and Winter () or compared to the features set size employed in the best systems employed in PAN CLEF 2014.…”

“…This strategy has the drawback of ignoring some stylistic features such as POS distribution, complex sentence construction measures, or other type‐token ratios. On the other hand, simpler text representation approaches have the advantage of simplicity, have proven to be efficient (Burrows, ; Hoover, ; Savoy, ), and can be understood by the final user. After an attribution has been proposed by the system, the final user may require a justification (e.g., in a court decision).…”

Text clustering: An application with the State of the Union addresses

“…In the current study, we use only the k most frequent word‐types (with k = 300) to compute the intertextual distance to reflect the stylistic information. The relative frequencies of these very frequent terms, mainly composed of functional words, tend to represent the fingerprint of each particular author and have been found effective in various authorship attribution studies (Burrows, ; Juola, ; Zhao & Zobel, ; Savoy, ).…”

Authorship of Pauline epistles revisited

“…This limit of 200 seems subjective. A recent study (Savoy, ) shows, however, that considering between 200 to 500 most frequent terms tends to produce the highest performance levels. Moreover, in the current case, some of the epistles are rather short (and some have fewer than 200 distinct words, for example, Philemon , 2 John ).…”