Japanese Dialogue Corpus of Information Navigation and Attentive Listening Annotated with Extended ISO-24617-2 Dialogue Act Tags

Japanese Dialogue Corpus of Information Navigation and Attentive Listening Annotated with Extended ISO-24617-2 Dialogue Act Tags Koichiro Yoshino1,3, Hiroki Tanaka1, Kyoshiro Sugiyama1, Makoto Kondo2 and Satoshi Nakamura1 1Graduate School of Information Science, Nara Institute of Science and Technology, 8916-5, Takayama-cho, Ikoma, 6300192, Japan 2Health Care Center, Nara Institute of Science and Technology, 8916-5, Takayama-cho, Ikoma, 6300192, Japan 3PRESTO, Japan Science and Technology Agency, 4-1-8 Hon-machi, Kawaguchi, Saitama 332-0012, Japan {koichiro, hiroki-tan, sugiyama.kyoshiro.sc7, s-nakamura}@is.naist.jp, makoto-k@hcc.naist.jp Abstract Large-scale dialogue data annotated with dialogue states is necessary to model a natural conversation with machines. However, large- scale conventional dialogue corpora are mainly built for specified tasks (e.g., task-oriented systems for restaurant or bus information navigation) with specially designed dialogue states. Text-chat based dialogue corpora have also been built due to the growth of social communication through the internet; however, most of them do not reflect dialogue behaviors in face-to-face conversation, including backchannelings or interruptions. In this paper, we try to build a corpus that covers a wider range of dialogue tasks than existing task-oriented systems or text-chat systems, by transcribing face-to-face dialogues held in natural conversational situations in tasks of information navigation and attentive listening. The corpus is recorded in Japanese and annotated with an extended ISO-24617-2 dialogue act tag-set, which is defined to see behaviors in natural conversation. The developed data can be used to build a dialogue model based on the ISO-24617-2 dialogue act tags. Keywords: Dialogue corpus, dialogue act, information navigation, attentive listening 1. Introduction Spoken dialogue systems have been modeled with ab- stracted classes, such as dialogue states, which are hand- crafted for the assumed task of the system (Dahl et al., 1994; Williams et al., 2013; Kim et al., 2016). However, handcrafting such dialogue states is costly, making it harder to build extensible dialogue systems for a variety of do- mains. Two solutions are widely used to avoid this prob- lem: building a multi-domain system (Gaˇsi´c et al., 2015; Papangelis and Stylianou, 2017) and building a system with more generalized classes (Yoshino et al., 2017; Keizer and Rieser, 2017). However, the approach of building a multi- domain system is an extension of using existing slot-value type dialogue states, and it is hard to apply this architec- ture to dialogue tasks that do not assume language under- standing methods based on slot-filling, such as information navigation (Yoshino and Kawahara, 2015) or attentive lis- tening (Yamaguchi et al., 2016; Lala et al., 2017; Tanaka et al., 2016). In comparison, general dialogue acts are more effectively used in these tasks because the approach of using them separates the functions and content of dia- logue and only focuses on modeling dialogue functions. In other words, the approach of using generalized dialogue act classes does not have the problem of exponentially increas- ing the number of dialogue states, which is caused by com- binations of dialogue content. ISO-24617-2 (Bunt et al., 2010; Bunt et al., 2012) is de- fined as a standard of dialogue act classes that is based on communicative functions of …

Content truncated to 500 words