Predicting Subjectivity in Multimodal Conversations

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 1348–1357, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Predicting Subjectivity in Multimodal Conversations Gabriel Murray and Giuseppe Carenini University of British Columbia Vancouver, Canada (gabrielm, carenini)@cs.ubc.ca Abstract In this research we aim to detect sub- jective sentences in multimodal conversa- tions. We introduce a novel technique wherein subjective patterns are learned from both labeled and unlabeled data, us- ing n-gram word sequences with vary- ing levels of lexical instantiation. Ap- plying this technique to meeting speech and email conversations, we gain signifi- cant improvement over state-of-the-art ap- proaches. Furthermore, we show that cou- pling the pattern-based approach with fea- tures that capture characteristics of gen- eral conversation structure yields addi- tional improvement. 1 Introduction Conversations are rich in subjectivity. Conversa- tion participants agree and disagree with one other, argue for and against various proposals, and gen- erally take turns expressing their private states. Being able to separate these subjective utterances from more objective utterances would greatly fa- cilitate the analysis, mining and summarization of a large number of conversations. Two of the most prevalent conversational me- dia are meetings and emails. Face-to-face meet- ings enable numerous people to exchange a large amount of information and opinions in a short pe- riod of time, while emails allow for concise ex- changes between potentially far-flung participants. Meetings and emails can also feed into one an- other, with face-to-face meetings occurring at reg- ular intervals and emails continuing the conver- sations in the interim. This poses several inter- esting questions, such as whether subjective utter- ances are more or less likely to be found in email exchanges compared with meetings, and whether the ratios of positive and negative subjective utter- ances differ between the two modalities. In this paper we describe a novel approach for predicting subjectivity, and test it in two sets of experiments on meetings and emails. Our ap- proach combines a new general purpose method for learning subjective patterns, with features that capture basic characteristics of conversation struc- ture across modalities. The subjective patterns are essentially n-gram sequences with varying levels of lexical instantiation, and we demonstrate how they can be learned from both labeled and un- labeled data. The conversation features capture structural characteristics of multimodal conversa- tions as well as participant information. We test our approach in two sets of experi- ments. The goal of the first set of experiments is to discriminate subjective from non-subjective utter- ances, comparing the novel approach to existing state-of-the-art techniques. In the second set of experiments, the goal is to discriminate positive- subjective and negative-subjective utterances, es- tablishing their polarity. In both sets of experi- ments, we assess the impact of features relating to conversation structure. 2 Related Research Raaijmakers et al. (2008) have approached the problem of detecting subjectivity in meeting speech by using a variety of multimodal features such as prosodic features, word n-grams, charac- ter n-grams and phoneme n-grams. For subjec- tivity detection, they found …

Content truncated to 500 words