Context-based Message Expansion for Disentanglement of Interleaved Text Conversations

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 200–208, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Context-based Message Expansion for Disentanglement of Interleaved Text Conversations Lidan Wang Computer Science Dept./UMIACS University of Maryland, College Park College Park, MD 20742 lidan@cs.umd.edu Douglas W. Oard College of Information Studies/UMIACS and HLT Center of Excellence University of Maryland, College Park College Park, MD 20742 oard@umd.edu Abstract Computational processing of text exchanged in interactive venues in which participants en- gage in simultaneous conversations can bene- fit from techniques for automatically grouping overlapping sequences of messages into sepa- rate conversations, a problem known as “dis- entanglement.” While previous methods ex- ploit both lexical and non-lexical information that exists in conversations for this task, the inter-dependency between the meaning of a message and its temporal and social contexts is largely ignored. Our approach exploits con- textual properties (both explicit and hidden) to probabilistically expand each message to provide a more accurate message representa- tion. Extensive experimental evaluations show our approach outperforms the best previously known technique. 1 Introduction Conversational media such as the text messages found in Internet Relay Chat presents both new op- portunities and new challenges. Among the chal- lenges are that individual messages are often quite short, for the reason that conversational participants are able to assemble the required context over the course of a conversation. A natural consequence of this is that many tasks that we would like to perform on conversational media (e.g., search, summariza- tion, or automated response) would benefit from re- assembly of individual messages into complete con- versations. This task has been studied extensively in the context of email (where it is often referred to as “threading”) (Yeh et al., 2006). The extensive meta- data associated with email and the relatively rich content of some email messages makes email some- what of a special case in the broad set of conversa- tion recovery tasks, however. At the opposite ex- treme, conversation “threading” in multi-party spo- ken interactions (e.g., meetings) would be a com- pelling application, but the word error rate of current automated transcription techniques somewhat limits access to the lexical evidence that we know is use- ful for this task. The recent interest in identifying individual conversations from online-discussions, a task that some refer to as “disentanglement,” there- fore seems to be something of a middle ground in the research space: computationally tractable, repre- sentative to some degree of a broader class of prob- lems, and directly useful as a pre-processing step for a range of important applications. One way to think of this task is as a clustering problem—we seek to partition the messages into a set of disjoint clusters, where each cluster represents a conversation among a set of participants on a topic. This formulation raises the natural question of how we should design a similarity measure. Since the messages are often too short to be meaningful by themselves, techniques based solely on lexical over- …

Content truncated to 500 words