Extraction of Novel Character Information from Synopses of Fantasy Novels in Japanese using Sequence Labeling

Extraction of Novel Character Information from Synopses of Fantasy Novels in Japanese using Sequence Labeling Yuji Oka Kagawa University Takamatsu, Kagawa, JAPAN s20g460@stu.kagawa-u.ac.jp Kazuaki Ando Kagawa University Takamatsu, Kagawa, JAPAN ando.kazuaki@kagawa-u.ac.jp Abstract In this paper, we propose a method of ex- tracting novel character information such as characters’ names, gender, age, occupations, and a part of relationships between characters from synopses of fantasy novels in Japanese using sequence labeling by comparing se- quence labeling models based on CRF (Con- ditional Random Fields) and deep learning with CRF. From the experimental results, we confirmed that the BiLSTM-CRF model with the information of part-of-speech of words has achieved the best performance, the pre- cision of 85.40%, the recall of 91.47%, and F1-measure of 88.30% for extracting charac- ters’ names. The BiLSTM-CRF model has achieved the best overall performance for ex- tracting all tags. 1 Introduction In recent years, the spread of electronic books has created an environment in which we can read novels easily. Moreover, websites or applications to post and publish new user-generated novels have also re- ceived a lot of attention all over the world. On the other hand, as the total number of published and posted novels increases, it will be difficult to find novels that suit a reader’s taste. We can search for novels by authors’ names, genres of novels, and key- words set independently by the author, and choose novels from the popularity ranking. However, we cannot search for the contents of novels based on personal preferences. In order to solve such problems, the following methods can be useful: (1) Search function using personal preferences for novels, (2) Synopsis gen- eration and presentation based on personal prefer- ences, (3) Generation and presentation of a correla- tion diagram of the characters in the novel. Personal preferences for novels can be divided into preferences for a story such as “surprise end- ing” and “happy ending”, and preferences for a novel character such as “handsome butler” and “dark hero”. As the first step to achieve three methods, we focus on the preferences for the novel characters. In this paper, we consider a method to extract informa- tion about novel characters. By extracting and orga- nizing the information about the characters from the novel text, it can be used not only for search targets, but also for generating synopses and correlation di- agrams. Although some of text data such as novel posting sites are available free of charge, most text data of novels published for commercial use must be used for a fee. In addition, when extracting informa- tion about characters from text of novels by super- vised machine learning, it is very costly to construct training data by annotating the whole text of novels. Therefore, we use “synopses” of novels as text for training data in this paper. A synopsis is attached to almost all novels regardless of whether it is pub- lished for commercial use, the length of text is short, and the cost of constructing learning data is …

Content truncated to 500 words