Stylometry Analysis of Literary Texts in Polish

In this work we compare different methods for deriving features for text representation in two stylometric tasks of gender and author recognition. The first group of methods uses the Bag-of-Words (BoW) approach, which represents the documents with vectors of frequencies of selected features occurring in the documents. We analyze features such as the most frequent 1000 lemmas, word forms, all lemmas, selected (content insensitive) lemmas, bigrams of grammatical classes and mixture of bigrams of grammatical classes, selected lemmas and punctuations. Moreover, the approach based on the recently proposed fastText algorithm (for vector based representation of text) is also applied. We evaluate these different approaches on two publicly available collections of Polish literary texts from late 19th- and early 20th-century: one consisting of 99 novels from 33 authors and the second one 888 novels from 58 authors. Our study suggests that depending on the corpora the best are the style features (grammatical bigrams) or semantic features (1000 lemmas extracted from the training set). We also noticed the importance of proper division of corpora into training and testing sets.