[4Rin1] Interactive 2

[4Rin1-82] Preliminary Experiment for Estimation Method of Emotion-Provoking Event Using BERT

〇Yoshiki Fukuda¹, Kohji Dohsaka¹, Masaki Ishii¹, Hidekatsu Ito¹ (1.Akita Prefectural University)

Keywords:text understanding, emotion estimation, BERT language representation model

Aiming at stimulate human affective communication, we are engaged in research on conversational robots that can not only recognize dialogue partner's emotion but also explain the reasoning for the judgment. For that purpose, the paper presents a system that, when a text and its writer's emotion is given, extracts a sentence describing the emotion provoking event. To develop the system, we first created an emotion provoking event database, which is comprised of 200,000 sentences describing emotion provoking events. The system calculates the similarity between an input sentence of the text and a sentence of the database by using the cosine similarity of the vector representations of two sentences, and it extracts an input sentence that can describe an emotion provoking event. Comparing sentence vectors based on BERT with those based on a conventional bag of words model, we show that, when sentence vectors based on BERT are used, the performance of extracting a sentence describing an emotional evoking event is improved. However, the error analysis indicates that we need to improve the quality and quantity of the database.