What Decisions Have You Made?: Automatic Decision Detection in Meeting Conversations

Proceedings of NAACL HLT 2007, pages 25–32, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics What Decisions Have You Made: Automatic Decision Detection in Conversational Speech Pei-Yun Hsueh School of Informatics University of Edinburgh Edinburgh EH9 8WL, UK p.hsueh@ed.ac.uk Johanna Moore School of Informatics University of Edinburgh Edinburgh EH9 8WL, UK J.Moore@ed.ac.uk Abstract This study addresses the problem of au- tomatically detecting decisions in conver- sational speech. We formulate the prob- lem as classifying decision-making units at two levels of granularity: dialogue acts and topic segments. We conduct an em- pirical analysis to determine the charac- teristic features of decision-making dia- logue acts, and train MaxEnt models using these features for the classification tasks. We find that models that combine lexi- cal, prosodic, contextual and topical fea- tures yield the best results on both tasks, achieving 72% and 86% precision, respec- tively. The study also provides a quantita- tive analysis of the relative importance of the feature types. 1 Introduction Making decisions is an important aspect of conver- sations in collaborative work. In the context of meet- ings, the proposed argumentative models, e.g., in Pallotta et al. (2005) and Rienks et al. (2005), have specified decisions as an essential outcome of meet- ings. Whittaker et al. (2005) have also described how reviewing decisions is critical to the re-use of meeting recordings. For example, a new engineer who just get assigned to a project will need to know what major decisions have been made in previous meetings. Unless all decisions are recorded in meet- ing minutes or annotated in the speech recordings, it is difficult to locate the decision points by the brows- ing and playback utilities alone. Banerjee and Rudnicky (2005) have shown that it is easier for users to retrieve the information they seek if the meeting record includes information about topic segmentation, speaker role, and meet- ing state (e.g., discussion, presentation, briefing). To assist users in identifying or revisiting decisions in meeting archives, our goal is to automatically iden- tify the dialogue acts and segments where decisions are made. Because reviewing decisions is indis- pensable in collaborative work, automatic decision detection is expected to lend support to computer- assisted meeting tracking and understanding (e.g., assisting in the fulfilment of the decisions made in the meetings) and the development of group infor- mation management applications (e.g., constructing group memory). 2 Related Work Spontaneous face-to-face dialogues in meetings vi- olate many assumptions made by techniques pre- viously developed for broadcast news (e.g., TDT and TRECVID), telephone conversations (e.g., Switchboard), and human-computer dialogues (e.g., DARPA Communicator). In order to develop techniques for understanding multiparty dialogues, smart meeting rooms have been built at several insti- tutes to record large corpora of meetings in natural contexts, including CMU (Waibel et al., 2001), LDC (Cieri et al., 2002), NIST (Garofolo et al., 2004), ICSI (Janin et al., 2003), and in the context of the IM2/M4 project (Marchand-Mailet, 2003). More recently, scenario-based meetings, in which partic- 25 ipants are assigned to different roles and …

Content truncated to 500 words