Multimodal Subjectivity Analysis of Multiparty Conversation

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 466–474, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Multimodal Subjectivity Analysis of Multiparty Conversation Stephan Raaijmakers TNO Information and Communication Technology Delft, The Netherlands stephan.raaijmakers@tno.nl Khiet Truong TNO Defense, Security and Safety Soesterberg, The Netherlands khiet.truong@tno.nl Theresa Wilson School of Informatics University of Edinburgh Edingburgh, UK twilson@inf.ed.ac.uk Abstract We investigate the combination of several sources of information for the purpose of sub- jectivity recognition and polarity classification in meetings. We focus on features from two modalities, transcribed words and acoustics, and we compare the performance of three dif- ferent textual representations: words, charac- ters, and phonemes. Our experiments show that character-level features outperform word- level features for these tasks, and that a care- ful fusion of all features yields the best perfor- mance. 1 1 Introduction Opinions, sentiments and other types of subjective content are an important part of any meeting. Meet- ing participants express pros and cons about ideas, they support or oppose decisions, and they make suggestions that may or may not be adopted. When recorded and archived, meetings become a part of the organizational knowledge, but their value is lim- ited by the ability of tools to search and summa- rize meeting content, including subjective content. While progress has been made on recognizing pri- marily objective meeting content, for example, in- formation about the topics that are discussed (Hsueh and Moore, 2006) and who is assigned to work on given tasks (Purver et al., 2006), there has been 1This work was supported by the Dutch BSIK-project Mul- timediaN, and the European IST Programme Project FP6- 0033812. This paper only reflects the authors’ views and fund- ing agencies are not liable for any use that may be made of the information contained herein. fairly little work specifically directed toward recog- nizing subjective content. In contrast, there has been a wealth of research over the past several years on automatic subjectiv- ity and sentiment analysis in text, including on-line media. Partly inspired by the rapid growth of so- cial media, such as blogs, as well as on-line news and reviews, researchers are now actively address- ing a wide variety of new tasks, ranging from blog mining (e.g., finding opinion leaders in an on-line community), to reputation management (e.g. find- ing negative opinions about a company on the web), to opinion-oriented summarization and question an- swering. Yet many challenges remain, including how best to represent and combine linguistic infor- mation for subjectivity analysis. With the additional modalities that are present when working with face- to-face spoken communication, these challenges are even more pronounced. The work in this paper focuses on two tasks: (1) recognizing subjective utterances and (2) discrimi- nating between positive and negative subjective ut- terances. An utterance may be subjective because the speaker is expressing an opinion, because the speaker is discussing someone else’s opinion, or be- cause the speaker is eliciting the opinion of someone else with a question. We approach the above …

Content truncated to 500 words