Invited Talk: Recognition and Understanding of Meetings

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 1–9, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Recognition and Understanding of Meetings Steve Renals Centre for Speech Technology Research, University of Edinburgh Informatics Forum, 10 Crichton Street, Edinburgh EH8 9AB, UK s.renals@ed.ac.uk homepages.inf.ed.ac.uk/srenals/ Abstract This paper is about interpreting human com- munication in meetings using audio, video and other signals. Automatic meeting recognition and understanding is extremely challenging, since communication in a meeting is sponta- neous and conversational, and involves mul- tiple speakers and multiple modalities. This leads to a number of significant research prob- lems in signal processing, in speech recog- nition, and in discourse interpretation, tak- ing account of both individual and group be- haviours. Addressing these problems requires an interdisciplinary effort. In this paper, I discuss the capture and annotation of multi- modal meeting recordings—resulting in the AMI meeting corpus—and how we have built on this to develop techniques and applications for the recognition and interpretation of meet- ings. 1 Introduction On the face of it, meetings do not seem to form a compelling research area. Although many people spend a substantial fraction of their time in meetings (e.g. the 1998 3M online survey at http://www. 3m.com/meetingnetwork/), for most people they are not the most enjoyable aspect of their work. However, for all the time that is spent in meet- ings, technological support for the meeting process is scant. Meeting records usually take the form of brief minutes, personal notes, and more recent use of collaborative web 2.0 software. Such records are labour intensive to produce—because they are man- ually created—and usually fail to capture much of the content of a meeting, for example the factors that led to a particular decision and the different sub- jective attitudes displayed by the meeting attendees. For all the time invested in meetings, very little of the wealth of information that is exchanged is ex- plicitly preserved. To preserve the information recorded in meet- ings, it is necessary to capture it. Obviously this involves recording the speech of the meeting partic- ipants. However, human communication is a mul- timodal activity with information being exchanged via gestures, handwritten diagrams, and numerous social signals. The creation of a rich meeting record involves the capture of data across several modal- ities. It is a key engineering challenge to capture such multimodal signals in a reliable, unobtrusive and flexible way, but the greater challenges arise from unlocking the multimodal recordings. If such recordings are not transcribed and indexed (at the least), then access merely corresponds to replay. And it is rare that people will have the time, or the inclination, to replay a meeting. There is a long and interesting thread of research which is concerned to better understand the dynam- ics of meetings and the way that groups function (Bales, 1951; McGrath, 1991; Stasser and Taylor, 1991). The types of analyses and studies carried out by these authors is still someway beyond what …

Content truncated to 500 words