Strategies for Mixed-Initiative Conversation Management using Question-Answer Pairs

Proceedings of COLING 2012: Technical Papers, pages 2821–2834, COLING 2012, Mumbai, December 2012. Strategies for Mixed-Initiative Conversation Management using Question-Answer Pairs Wilson Wong, Lawrence Cavedon, John Thangarajah, Lin Padgham RMIT University, Melbourne, Australia {wilson.wong,lawrence.cavedon,john.thangarajah,lin.padgham}@rmit.edu.au ABSTRACT One of the biggest bottlenecks for conversational systems is large-scale provision of suitable content. In this paper, we present the use of content mined from online question-and-answer forums to automatically construct system utterances. Although this content is mined in the form of question-answer pairs, our system is able to use it to formulate utterances that drive a conversation, not just for answering user questions as has been done in previous work. We use a collection of strategies that specify how and when the question-answer pairs can be used and augmented with a small number of generic hand-crafted text snippets to generate natural and coherent system utterances. Our experiments involving 11 human participants demonstrated that this approach can indeed produce relatively natural and coherent interaction. KEYWORDS: conversational system; question-answer pairs; conversational strategies. 1 Introduction A major bottleneck for any kind of conversational system is provision of sufficient content, and if one wishes to avoid repetition and to cover the many possible user inputs, there must be a substantial amount of content. In the case of chatbots such as ALICE (Wallace, 2009), the content is custom-crafted to be as generic and deflective as possible to cover the needs of open-ended conversations. This approach, which reduces the amount of content that the system needs despite the relatively broad range of user inputs, is essentially only capable of content-free, small talk. This kind of conversational ability is, however, inadequate for virtual characters that need to contribute relevant content in their conversations. In systems that require domain-specific content, for example about health in a virtual nurse (Bickmore et al., 2009) or about children-related topics in an intelligent toy (Chen et al., 2011), the custom- crafting of content to cover the depth and breadth of the domain as well as to keep it current becomes an impractical task. This work reports on our use of an abundant type of data, sourced from the Web, to equip a conversational agent with the content it needs to engage users in a coherent and natural way over a wide range of domains. To overcome the bottleneck of custom- or hand-crafting conversational content, researchers have recently investigated mining online resources such as Web pages returned by search engines (e.g., Yahoo! (Shibata et al., 2009)). The problem with this approach, however, is that most content on the Web is not suitable for conversational agents, for various reasons. The verbose, non-colloquial and monologue nature of much of Web text is not a good match for the characteristics of human-human dialogue. However, there are some other sources on the Web that have more potential for this purpose, such as human-written comments on news websites (e.g., NYTimes.com (Marge et al., 2010)) and online forums (e.g., RottenTomatoes.com 2821 (Huang et al., 2007), 2ch.net (Inoue et al., 2011)). We use text …

Content truncated to 500 words