A Corpus of Simulated Counselling Sessions with Dialog Act Annotation

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5723–5730 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5723 A Corpus of Simulated Counselling Sessions with Dialog Act Annotation John S. Y. Lee1, Haley H. M. Fong1, L. S. Judy Wong1, Chun Chung Mak1, Chi Hin Yip1, C. W. Larry Ng2 1Department of Linguistics and Translation, City University of Hong Kong 2Student Development Services, City University of Hong Kong Hong Kong SAR, China jsylee@cityu.edu.hk, heimfong3-c@my.cityu.edu.hk, judy99@hotmail.com, chuncmak@cityu.edu.hk, ychihin@gmail.com, larryng@cityu.edu.hk Abstract We present a corpus of simulated counselling sessions consisting of speech- and text-based dialogs in Cantonese. Consisting of 152K Chinese characters, the corpus labels the dialog act of both client and counsellor utterances, segments each dialog into stages, and identifies the forward and backward links in the dialog. We analyze the distribution of client and counsellor communicative intentions in the various stages, and discuss significant patterns of the dialog flow. Keywords: dialog act, dialog corpus, counselling, Cantonese 1. Introduction With the rapid growth in demand for counselling in re- cent years, corpus-based analysis of counselling con- versations has seen increasing interest (Althoff et al., 2016; P´erez-Rosas et al., 2017; Inoue et al., 2012). These corpora can be useful both for practitioners and for various tasks in chatbot development. First, they can lead to better understanding of the characteris- tics of successful counselling, for example the lin- guistic markers of empathy (P´erez-Rosas et al., 2017) and the use of forward- and backward-oriented lan- guage (Zhang and Danescu-Niculescu-Mizil, 2020). These corpora can also help identify the client’s com- municative intention, or dialog act (DA), and thereby determine the most appropriate DA for the counsel- lor’s reply. Research has shown that the communica- tive intention of the interlocutor can inform response retrieval (Li et al., 2017). DA has been extensively studied in the general domain (Amanova et al., 2016), with support from large-scale resources to facilitate DA modeling (Stolcke et al., 2000; Kim et al., 2010a; Cadilhac et al., 2013). Supervised models have also been implemented to classify the client DA and fore- cast the DA of the counsellor’s response (Cao et al., 2019; Park et al., 2019). We present a corpus that is designed to support the development of counsellor chatbots in Cantonese, the most widely spoken variety of Chinese (Matthews and Yip, 2011). Similar to corpora developed by Stiles et al. (1988) and Cao et al. (2019), it annotates both client and counsellor DAs, hence facilitating analysis of the communicative intentions of both parties. To the best of our knowledge, this is the first publicly available corpus1 of counselling sessions in Cantonese. 1https://github.com/CantoneseCounsellorChatbot Subcorpus Speech Text # dialogs 29 57 # turns per dialog 72.03 53.68 Counsellor turn length (# chars) 35.44 30.71 Client turn length (# chars) 29.63 24.66 Table 1: Corpus statistics for the speech-based (Sec- tion 2.1) and text-based dialogs (Section 2.2) A significant feature of the corpus is the annotation of not only the DAs, but also forward …

Content truncated to 500 words