Which parts of speech are more susceptible to topic variations: a case of Japanese authorship analysis

要旨

Abstract We examined which parts of speech are susceptible to topic variations in Japanese authorship analysis. (1) the 4,574 texts written by 2287 authors extracted from “Yahoo! Blog” were categorized into six topics using latent Dirichlet allocation among topic models. (2) Four groups of pair texts were created using hierarchical cluster analysis and multidimensional scaling, based on topic consistency (same or different topics) and a combination of authorship between paired texts (same or different authorships). (3) The degree of the effect arising from topic differences was examined using Bayes models and effect sizes by focusing on phrase patterns, which were divided into phrase units and masked words with part-of-speech (POS) tags, except for target POS words within each phrase. This study concluded that topic effects exist across both content and function words. Additionally, noun words were likely to be highly influenced by topic effects and postpositional particles showed the lowest values.

主題