A Comparison of Latent Variable Models For Conversation Analysis

With the evolution of online communication methods, conversations are increasingly handled via email, internet forums and other such methods. In this paper, we attempt to model lexical information in a context sensitive manner, encoding our belief that the use of language depends on the participants in the conversation. We model the discourse as a combination of the speaker, the addressee and other participants in the conversation as well as a context specific language model. In order to do this, we introduce a novel method based on an HMM with an exponential state space to capture speaker-addressee context. We also study the performance of topic modeling frameworks in conversational settings. We evaluate the models on the tasks of identifying the set of people present in any conversation, as well as identifying the speaker for every utterance in the conversation, and they show significant improvement over the baseline models.