System Utterance Generation by Label Propagation over Association Graph of Words and Utterance Patterns for Open-Domain Dialogue Systems

System Utterance Generation by Label Propagation over Association Graph of Words and Utterance Patterns for Open-Domain Dialogue Systems Hiroshi Tsukahara Kei Uchiumi Denso IT Laboratory, Inc. Shibuya Cross Tower 28F, 2-15-1 Shibuya, Shibuya-ku, Tokyo, Japan Abstract A novel graph-based utterance generation method for open-domain dialogue systems is proposed in this paper. After an association graph of words and utterance patterns from a dialogue corpus is constructed, a label propa- gation algorithm is used for generating system utterances from the words and utterance pat- terns in the association graph that are found to strongly correlate with the words and ut- terance patterns that appeared in previous user utterances. We also propose a crowdsourcing framework for collecting annotated chat data so that we can implement our method in a cost effective manner. Crowdsourcing is also used for conducting subjective evaluations and the results will show that the proposed method can not only provide interesting and informative responses but it also can appropriately expand the topics by comparing them to a well-known chat system in Japanese. 1 Introduction Chatting plays a lot of important roles in human communications for naturally exchanging diverse information, facilitating collaborative tasks, or even enhancing the quality of the conversations them- selves. For dialogue systems as well, the function- ality of being able to create chats is considered to have a significant importance regardless of whether task-oriented or non-task-oriented. There are cur- rently many types of smart devices in our daily life and most of them have spoken dialogue interfaces, although they are basically limited to question- answering. However, there are cases where people do not always have the clear intent on searching for something but they just want to know whether there is anything interesting they should know. In such cases, if the systems could offer a chats function in- stead, people may be able to make such unconscious or potential intentions clear by themselves through chats with these systems. However, it is quite challenging for dialogue sys- tems to automatically generate chat responses be- cause of the wide variety of topics in user utter- ances. In ordinary dialogue systems, i.e., rule-based systems, a very large number of hand-crafted rules and utterance patterns, or templates, would need to be prepared for extending the coverage of topics they can handle. However, this would be a very formidable task both to create them and to maintain them while keeping them up to date. Thus, a data- driven approach that makes use of the huge amount of conversational resources currently on the web, such as microblogs or social network media, as cor- pora have been recently investigated (Shibata et al., 2009; Sugiyama et al., 2013). These corpora contain a large number of sentences that cover a wide range of topics, but there are many noisy sentences that do not contain meaningful content themselves. An- other issue with this approach is that it basically se- lects sentences that are similar to the user utterances on the surface-level. Thus, the generated responses tend …

Content truncated to 500 words