Resolving Personal Names in Email Using Context Expansion

Proceedings of ACL-08: HLT, pages 941–949, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Resolving Personal Names in Email Using Context Expansion Tamer Elsayed,∗Douglas W. Oard,† and Galileo Namata∗ Human Language Technology Center of Excellence and UMIACS Laboratory for Computational Linguistics and Information Processing (CLIP) University of Maryland, College Park, MD 20742 {telsayed, oard, gnamata}@umd.edu Abstract This paper describes a computational ap- proach to resolving the true referent of a named mention of a person in the body of an email. A generative model of mention gener- ation is used to guide mention resolution. Re- sults on three relatively small collections indi- cate that the accuracy of this approach com- pares favorably to the best known techniques, and results on the full CMU Enron collection indicate that it scales well to larger collections. 1 Introduction The increasing prevalence of informal text from which a dialog structure can be reconstructed (e.g., email or instant messaging), raises new challenges if we are to help users make sense of this cacophony. Large collections offer greater scope for assembling evidence to help with that task, but they pose addi- tional challenges as well. With well over 100,000 unique email addresses in the CMU version of the Enron collection (Klimt and Yang, 2004), common names (e.g., John) might easily refer to any one of several hundred people. In this paper, we associate named mentions in unstructured text (i.e., the body of an email and/or the subject line) to modeled iden- tities. We see at least two direct applications for this work: (1) helping searchers who are unfamiliar with the contents of an email collection (e.g., historians or lawyers) better understand the context of emails that they find, and (2) augmenting more typical social networks (based on senders and recipients) with ad- ditional links based on references found in unstruc- tured text. Most approaches to resolving identity can be de- composed into four sub-problems: (1) finding a ref- erence that requires resolution, (2) identifying can- didates, (3) assembling evidence, and (4) choosing ∗Department of Computer Science †College of Information Studies among the candidates based on the evidence. For the work reported in this paper, we rely on the user to designate references requiring resolution (which we model as a predetermined set of mention-queries for which the correct referent is known). Candidate identification is a computational expedient that per- mits the evidence assembly effort to be efficiently focused; we use only simple techniques for that task. Our principal contributions are the approaches we take to evidence generation (leveraging three ways of linking to other emails where evidence might be found: reply chains, social interaction, and topical similarity) and our approach to choosing among can- didates (based on a generative model of reference production). We evaluate the effectiveness of our approach on four collections, three of which have previously reported results for comparison, and one that is considerably larger than the others. The remainder of this paper is as follows. Sec- tion 2 surveys prior work. …

Content truncated to 500 words