Exploiting Conversation Structure in Unsupervised Topic Segmentation for Emails

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 388–398, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Exploiting Conversation Structure in Unsupervised Topic Segmentation for Emails Shafiq Joty and Giuseppe Carenini and Gabriel Murray and Raymond T. Ng {rjoty, carenini, gabrielm, rng}@cs.ubc.ca Department of Computer Science University of British Columbia Vancouver, BC, V6T 1Z4, Canada Abstract This work concerns automatic topic segmen- tation of email conversations. We present a corpus of email threads manually annotated with topics, and evaluate annotator reliabil- ity. To our knowledge, this is the first such email corpus. We show how the existing topic segmentation models (i.e., Lexical Chain Seg- menter (LCSeg) and Latent Dirichlet Alloca- tion (LDA)) which are solely based on lex- ical information, can be applied to emails. By pointing out where these methods fail and what any desired model should consider, we propose two novel extensions of the models that not only use lexical information but also exploit finer level conversation structure in a principled way. Empirical evaluation shows that LCSeg is a better model than LDA for segmenting an email thread into topical clus- ters and incorporating conversation structure into these models improves the performance significantly. 1 Introduction With the ever increasing popularity of emails and web technologies, it is very common for people to discuss issues, events, agendas or tasks by email. Effective processing of the email contents can be of great strategic value. In this paper, we study the problem of topic segmentation for emails, i.e., grouping the sentences of an email thread into a set of coherent topical clusters. Adapting the stan- dard definition of topic (Galley et al., 2003) to con- versations/emails, we consider a topic is something about which the participant(s) discuss or argue or express their opinions. For example, in the email thread shown in Figure 1, according to the major- ity of our annotators, participants discuss three top- ics (e.g., ‘telecon cancellation’, ‘TAG document’, and ‘responding to I18N’). Multiple topics seem to occur naturally in social interactions, whether syn- chronous (e.g., chats, meetings) or asynchronous (e.g., emails, blogs) conversations. In multi-party chat (Elsner and Charniak, 2008) report an average of 2.75 discussions active at a time. In our email cor- pus, we found an average of 2.5 topics per thread. Topic segmentation is often considered a pre- requisite for other higher-level conversation analy- sis and applications of the extracted structure are broad, encompassing: summarization (Harabagiu and Lacatusu, 2005), information extraction and or- dering (Allan, 2002), information retrieval (Dias et al., 2007), and intelligent user interfaces (Dredze et al., 2008). While extensive research has been con- ducted in topic segmentation for monologues (e.g., (Malioutov and Barzilay, 2006), (Choi et al., 2001)) and synchronous dialogs (e.g., (Galley et al., 2003), (Hsueh et al., 2006)), none has studied the problem of segmenting asynchronous multi-party conversa- tions (e.g., email). Therefore, there is no reliable an- notation scheme, no standard corpus, and no agreed- upon metrics available. Also, it is our …

Content truncated to 500 words