DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 986–995, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset †Yanran Li∗, †,§Hui Su∗, ‡Xiaoyu Shen, †Wenjie Li, †Ziqiang Cao, §Shuzi Niu †Department of Computing, The Hong Kong Polytechnic University, Hong Kong §Institute of Software, Chinese Academy of Science, China ‡Spoken Language Systems (LSV), Saarland University, Germany Abstract We develop a high-quality multi-turn dia- log dataset, DailyDialog, which is intrigu- ing in several aspects. The language is human-written and less noisy. The dia- logues in the dataset reflect our daily com- munication way and cover various topics about our daily life. We also manually la- bel the developed dataset with communi- cation intention and emotion information. Then, we evaluate existing approaches on DailyDialog dataset and hope it benefit the research field of dialog systems1. 1 Introduction Developing intelligent chatbots and dialog systems is of great significance to both commercial and aca- demic camps. A good conversational agent enables enterprises to provide automatic customer services and thus reduce human labor costs. For academia, it is challenging yet appealing to build up such an intelligent chatbot which involves a series of high-level natural language processing techniques, such as understanding the underlying semantics of user input utterance, and generating coherent and meaningful responses. However, the training datasets for this research area are still deficient. Traditional dialogue sys- tems are often trained with domain-specific spoken dialogue datasets (Ringger et al., 1996; Petukhova et al., 2014), which are often small-scale and ori- ented to complete a specific task. More recent work feed their conversational models with open-domain datasets. Switchboard (Godfrey et al., 1992) and OpenSubtitles (J¨org Tiedemann, 2009) datasets ∗Authors contributed equally. Correspondence should be sent to Y. Li (csyli@comp.polyu.edu.hk). 1The dataset is available on http://yanran.li/ dailydialog A: I’m worried about something. B: What’s that? A: Well, I have to drive to school for a meeting this morning, and I’m going to end up getting stuck in rush-hour traffic. B: That’s annoying, but nothing to worry about. Just breathe deeply when you feel yourself getting upset. A: Ok, I’ll try that. B: Is there anything else bothering you? A: Just one more thing. A school called me this morning to see if I could teach a few classes this weekend and I don’t know what to do. B: Do you have any other plans this weekend? A: I’m supposed to work on a paper that’d due on Monday. B: Try not to take on more than you can handle. A: You’re right. I probably should just work on my paper. Thanks! Figure 1: An example in DailyDialog dataset. Some text is shortened for space. Best viewed in color. comprise approximately 150 turns in a “conversa- tion” and thus are too disperse to capture the main topic. Twitter Dialog Corpus (Ritter et al., 2011) and Chinese Weibo dataset (Wang et al., 2013) are comprised of posts and replies on social networks, which are noisy, informal …

Content truncated to 500 words