Real-time decision detection in multi-party dialogue

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1133–1141, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Real-time decision detection in multi-party dialogue Matthew Frampton, Jia Huang, Trung Huu Bui and Stanley Peters Center for the Study of Language and Information Stanford University Stanford, CA, 94305, USA {frampton|jiahuang|thbui|peters}@stanford.edu Abstract We describe a process for automatically detecting decision-making sub-dialogues in multi-party, human-human meetings in real-time. Our basic approach to decision detection involves distinguishing between different utterance types based on the roles that they play in the formulation of a de- cision. In this paper, we describe how this approach can be implemented in real-time, and show that the resulting system’s per- formance compares well with other detec- tors, including an off-line version. 1 Introduction In collaborative and organized work environ- ments, people share information and make de- cisions through multi-party conversations, com- monly referred to as meetings. The demand for automatic methods that process, understand and summarize information contained in audio and video recordings of meetings is growing rapidly, as evidenced by on-going projects which are fo- cused on this goal, (Waibel et al., 2003; Janin et al., 2004). Our research is part of a general effort to develop a system that can automatically extract and summarize information such as conversational topics, action items, and decisions. This paper concerns the development of a real- time decision detector — a system which can de- tect and summarize decisions as they are made during a meeting. Such a system could provide a summary of all of the decisions which have been made up until the current point in the meeting, and this is something which we expect will help users to enjoy more productive meetings. Cer- tainly, good decision-making relies on access to relevant information, and decisions made earlier in a meeting often have a bearing on the current topic of discussion, and so form part of this rele- vant information. However, in a long and winding meeting, participants might not have these earlier decisions at the forefront of their minds, and so an accurate and succinct reminder, as provided by a real-time decision detector, could potentially be very useful. A record of earlier decisions could also help users to identify outstanding issues for discussion, and to therefore make better use of the remainder of the meeting. Our approach to decision detection uses an an- notation scheme which distinguishes between dif- ferent types of utterance based on the roles which they play in the decision-making process. Such a scheme facilitates the detection of decision discus- sions (Fern´andez et al., 2008), and by indicating which utterances contain particular types of infor- mation, it also aids their summarization. To auto- matically detect decision discussions, we use what we refer to as hierarchical classification. Here, in- dependent binary sub-classifiers detect the differ- ent decision dialogue acts, and then based on the sub-classifier hypotheses, a super-classifier deter- mines which dialogue regions are decision discus- sions. In this paper then, we address the …

Content truncated to 500 words