A Just-In-Time Keyword Extraction from Meeting Transcripts

Proceedings of NAACL-HLT 2013, pages 888–896, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics A Just-In-Time Keyword Extraction from Meeting Transcripts Hyun-Je Song Junho Go Seong-Bae Park School of Computer Science and EngineeringSe-Young Park Kyungpook National University Daegu, Korea {hjsong,jhgo,sbpark,sypark}@sejong.knu.ac.kr Abstract In a meeting, it is often desirable to extract keywords from each utterance as soon as it is spoken. Thus, this paper proposes a just-in- time keyword extraction from meeting tran- scripts. The proposed method considers two major factors that make it different from key- word extraction from normal texts. The first factor is the temporal history of preceding ut- terances that grants higher importance to re- cent utterances than old ones, and the sec- ond is topic relevance that forces only the pre- ceding utterances relevant to the current utter- ance to be considered in keyword extraction. Our experiments on two data sets in English and Korean show that the consideration of the factors results in performance improvement in keyword extraction from meeting transcripts. 1 Introduction A meeting is generally accomplished by a number of participants and a wide range of subjects are dis- cussed. Therefore, it would be helpful to meeting participants to provide them with some additional information related to the current subject. For in- stance, assume that a participant is discussing a spe- cific topic with other participants at a meeting. The summary of previous meetings on the topic is then one of the most important resources for her discus- sion. In order to provide information on a topic to par- ticipants, keywords should be first generated for the topic since keywords are often representatives of a topic. A number of techniques have been proposed for automatic keyword extraction (Frank et al., 1999; Turney, 2000; Mihalcea and Tarau, 2004; Wan et al., 2007), and they are designed to extract keywords from a written document. However, they are not suitable for meeting transcripts. In a meeting, it is often desirable to extract keywords at the time at which a new utterance is made for just-in-time ser- vice of additional information. Otherwise, the ex- tracted keywords become just the important words at the end of the meeting. Two key factors for just-in-time keyword extrac- tion from meeting transcripts are time of preceding utterances and topic of current utterance. First, cur- rent utterance is affected by temporal history of pre- ceding utterances. That is, when a new utterance is made it is likely to be related more closely with latest utterances than old ones. Second, the preced- ing utterances which carry similar topics to current utterance are more important than irrelevant utter- ances. Since a meeting consists of several topics, the utterances that have nothing to do with current utterance are inappropriate as a history of the cur- rent utterance. This paper proposes a graph-based keyword ex- traction to reflect these factors. The proposed method represents an utterance as a graph of which nodes are candidate keywords. The preceding utter- ances are also expressed as a …

Content truncated to 500 words