Relevance of named entities in authorship attribution

Abstract

Named entities (NE) are words that refer to names of people, locations, organization, etc. NE are present in every kind of documents: e-mails, letters, essays, novels, poems. Automatic detection of these words is very important task in natural language processing. Sometimes, NE are used in authorship attribution studies as a stylometric feature. The goal of this paper is to evaluate the effect of the presence of NE in texts for the authorship attribution task: are we really detecting the style of an author or are we just discovering the appearance of the same NE. We used the corpus that consists of 91 novels of 7 authors of XVIII century. These authors spoke and wrote English, their native language. All novels belong to fiction genre. The used stylometric features were character n-grams, word n-gram and n-gram of POS tags of various sizes (2-grams, 3-grams, etc.). Five novels were selected for each author, these novels contain between 4 and 7% of the NE. All novels were divided into blocks, each block contains 10,000 terms. Two kinds of experiment were conducted: automatic classification of blocks containing NE and of the same blocks without NE. In some cases, we use only the most frequent n-grams (500, 2,000 and 4,000 n-grams). Three machine learning algorithms were used for classification task: NB, SVM (SMO) and J48. The results show that as a tendency the presence of the NE helps to classify (improvements from 5% to 20%), but there are specific authors when NE do not help and even make the classification worse (about 10% of experimental data).