Event

Speaker Role Detection in Meetings using Social Network

Neha Priyadarshini Garg

EPFL and ICSI

Friday, February 08, 2008
2:00 p.m.

Meetings are part of everyone's day to day life. Usually, minutes are taken to record what took place during the meeting. However, this can be avoided by recording the meetings and then later navigating through the content of meeting. Previous work has shown that, due to the serial nature of speech, navigation of such data is difficult, and structural information is useful for easier navigation. Users can benefit from higher level structural information, such as who lead the discussion? Who implemented the decisions etc. Identification of speaker roles can be helpful in providing such high level information about meetings. Previous work on speaker role detection either mainly focused on using interaction patterns or lexical information for broadcast news or conversations.

In this work, we compare various features that can be used for detecting speaker roles in the project-oriented and naturally occurring meetings (such as in the AMI and ICSI meetings corpora). In particular, we analyze how much the knowledge of social networks present in the meeting is helpful to detect the speaker roles along with textual information. Our results show that social network analysis measures and lexical features are both useful for speaker role detection in meetings and their combination results in the best performance.