Literary Analysis Based on Distributional Semantics

Word embeddings demonstrate state-of-the-art performance as a set of models in distributional semantics. A number of research studies have shown that these models trained on huge text corpora handle various semantic tasks. This paper explores the suitability of word embeddings in the specific task domain of social network extraction from literary fiction with comparably small corpus sizes. The models are firstly trained on popular fantasy novel book series, namely “A Song of Ice and Fire” and “Harry Potter”, and further their performance is evaluated. The extracted social networks are compared to the “gold standard” obtained by crowdsourcing. This study focuses on text preprocessing extending frequency of character contexts via different state-of-the-art coreference resolution algorithms, as well as by replacing various character aliases with some standard form.