Author Clustering with and Without Topical Features

Abstract

Typically, the task of authorship attribution has been solved using supervised machine learning methods. It is only recently that unsupervised methods have been applied to authorship attribution, namely author clustering. Clustering could be useful in realistic scenario as it represents natural grouping of documents without a priori authorship information, although the problem of feature selection remains unsolved. That is particularly true for a cross-domain scenario. Studies have shown that in cross-domain settings some domain-specific text features cause noise in authorship attribution. In the current work we introduce a modification of unmasking technique aimed at selecting and removing the features most influenced by topic change. We apply the proposed technique to identify topical features and assess the quality of author clustering with different feature sets in a real-world dataset of forum texts in Russian. The main assumption is that the topical features result in topic-based text instead authorship-based clustering, and removing them could increase the performance of document clustering against authorship ground truth. We test this consideration by first clustering cross-topic documents with state-of-the-art authorship attribution features. Second, we remove the most significant topical features, and cluster texts with resulting feature set. Both clustering results are evaluated against ground truth authorship. The results demonstrate that the described approach of removing some topical features increases author clustering performance, however one should be cautious with the number of removed features.