JDCFC: A Japanese Dialogue Corpus with Feature Changes

JDCFC: A Japanese Dialogue Corpus with Feature Changes Tetsuaki Nakamura†, Daisuke Kawahara†‡ †Kyoto University, ‡JST PRESTO Yoshida-Honmachi, Sakyo-ku, Kyoto, Japan tnakamura@nlp.ist.i.kyoto-u.ac.jp, dk@i.kyoto-u.ac.jp Abstract In recent years, the importance of dialogue understanding systems has been increasing. However, it is difficult for computers to deeply understand our daily conversations because we frequently use emotional expressions in conversations. This is partially because there are no large-scale corpora focusing on the detailed relationships between emotions and utterances. In this paper, we propose a dialogue corpus constructed based on our knowledge base, called the Japanese Feature Change Knowledge Base (JFCKB). In JFCKB and the proposed corpus, the feature changes (mainly emotions) of arguments in event sentences (or utterances) and those of the event sentence recognizers (or utterance recognizers) are associated with the event sentences (or utterances). The feature change information of arguments in utterances and those of the utterance recognizers, replies to the utterances, and the reasonableness of the replies were gathered through crowdsourcing tasks. We conducted an experiment to investigate whether a machine learning method can recognize the reasonableness of a given conversation. Experimental result suggested the usefulness of our proposed corpus. Keywords: emotion, commonsense knowledge, dialogue corpus, crowdsourcing, neural network, LSTM 1. Introduction In recent years, the importance of dialogue understand- ing systems has been increasing because interactive inter- faces handling a natural language such as smart speakers have become popular. However, it is difficult for computer programs to understand our daily conversations because we frequently use emotional expressions in conversations. This is partially because there are no large-scale corpora fo- cusing on the detailed relationships between emotions and utterances. Many dialogue corpora have been developed because they are essential language resources needed to train and eval- uate machine learning methods. For instance, the Di- alog State Tracking Challenge (DSTC) dataset is used to estimate a user’s goal in a spoken dialog system (Kim et al., 2016). While the DSTC corpus is made from manually transcribed Skype dialogues, there are corpora that consist of conversations extracted from SNS websites (Ritter et al., 2010; Ritter et al., 2011; Sordoni et al., 2015; Shang et al., 2015). There is also a corpus based on a col- lection of logs extracted from Ubuntu-related chat rooms that is mainly composed of technical support conversa- tions (Lowe et al., 2015; Lowe et al., 2016). The Dialogue Breakdown Detection Challenge database is a corpus used to detect incorrect replies generated by dialogue systems (Higashinaka et al., 2016). Although these corpora are very useful resources for understanding actual human-human di- alogue or human-machine dialogue, it is difficult to under- stand a speaker’s/replier’s motivations because such cor- pora do not record a speaker’s/replier’s inner state (in par- ticular, his/her emotions). Even if such corpora include some keywords as clues for inferring a speaker’s/replier’s inner state, it is necessary to develop a method to extract inner state information from the corpora, which are com- posed of raw text. Dialogue corpora that include various feature changes of arguments in utterances and the reactions …

Content truncated to 500 words