Speaker Identification of Quotes in Japanese Novels based on Gender Classification Model by BERT

Speaker Identification of Quotes in Japanese Novels based on Gender Classification Model by BERT Yuki Zenimoto, Takehito Utsuro Degree Programs in Systems and Information Engineering, Graduate School of Science and Technology, University of Tsukuba, 1-1-1, Tennodai, Tsukuba, Ibaraki, 305-8573, Japan s2220753 @ s.tsukuba.ac.jp, utsuro @ .tsukuba.ac.jp Abstract In the Japanese language, particles and auxil- iary verbs in utterances tend to vary depend- ing on the speaker’s gender. Thus, the lin- guistic expressions within the utterance are an important hint in identifying the speaker. This research proposes a method for identi- fying the speaker in novels using linguistic expressions within the utterances that reflect the gender. We constructed a dataset of utter- ances with gender-specific linguistic expres- sions by automatically collecting utterances from novels that contained the first-person pronouns “俺 (ore)” or “私 (watashi)”, con- sidering that “俺 (ore)” is primarily used by males and “私 (watashi)” is primarily used by females. We fine-tuned a BERT (Devlin et al., 2019)-based gender-specific language model that classifies the gender of the speaker of a given utterance using this dataset as the train- ing data. The fine-tuned gender classification model achieved an accuracy of 87.9% when evaluating the utterances of two main charac- ters in a Japanese romance novel, demonstrat- ing that this gender classification model is ef- fective in speaker identification. 1 Introduction Novels are valuable resources for enriching dialogue systems. Various characters converse with each other in novels, and their behaviors are described in the narratives. If such information is appropriately extracted, it can be used in dialogue systems and var- ious research such as personality analysis. Dialogue systems are required to respond appropriately to the users’ utterances, and it is also expected to have a specific personality. To train dialogue models, it is necessary to collect dialogue data from web ser- vices such as Twitter and Reddit (Serban et al., 2015; Mazar´e et al., 2018). However, these dialogue data are insufficient for training dialogue models ca- pable of acting as if they have specific personali- ties. This problem is simply because these dialogue data consist of a mixture of utterances of numerous personalities and are difficult to utilize as a source for building dialogue models with specific person- alities. Considering the preceding discussion, it is required to prepare dialogue data that are accompa- nied by profiles (Zhang et al., 2018; Sugiyama et al., 2021). Crowdsourcing is commonly used to collect such dialogue data with profile information. How- ever, crowdsourcing large scale dialogue data with profile information is extremely expensive. Considering this limitation, this paper introduced an approach for gathering dialogue data from nov- els, examining the task of speaker identification in novels to generate persona data. Previous work on speaker identification in English novels (O’Keefe et al., 2012; Muzny et al., 2017) relied heavily on nar- rative elements surrounding the target utterance and the utterances before and after it. However, the tar- get utterance itself was not effectively used. Here, it is important to note that, in the Japanese language, particles …

Content truncated to 500 words