IDENTITY RESOLUTION IN EMAIL COLLECTIONS

Abstract

Access to historically significant email collections poses challenges that arise less often in personal collections. Most notably, people exploring a large collection of emails, in which they were not sending or receiving, may not be very familiar with the discussions that exist in this collection. They would not only need to focus on understanding the topical content of those discussions, but would also find it useful to understand who the people sending, receiving, or mentioned in these discussions were. In this dissertation, the problem of resolving personal identity in the context of large email collections is tackled. In such collections, a common name (e.g., John) might easily refer to any one of several hundred people; when one of these people was mentioned in an email, the question then arises: "who is that John?'' To "resolve identity'' of people in an email collection, two problems need to be solved: (1) modeling the identity of the participants in that collection, and (2) resolving name-mentions (that appeared in the body of the messages) to these identities. To tackle the first problem, a simple computational model of identity, that is built on extracting unambiguous references (e.g., full names from headers, or nicknames from free-text signatures) to people from the whole collection, is presented. To tackle the second problem, a generative probabilistic approach that leverages the model of identity to resolve mentions is presented. The approach is motivated by intuitions about the way people might refer to others in an email; it expands the context surrounding a mention in four directions: the message where the mention was observed, the thread that includes that message, topically-related messages, and messages sent or received by the original communicating parties. It relies on less ambiguous references (e.g., email addresses or full names) that are observed in some context of a given mention to rank potential referents of that mention. In order to jointly resolve all mentions in the collection, a parallel implementation is presented using the MapReduce distributed-programming framework. The implementation decomposes the structure of the resolution process into subcomponents that fit the MapReduce task model well. At the heart of that implementation, a parallel algorithm for efficient computation of pairwise document similarity in large collections is proposed as a general solution that can be used for scalable context expansion of all mentions and other applications as well. The resolution approach compares favorably with previously-reported techniques on small test collections (sets of mention-queries that were manually resolved beforehand) that were used to evaluate the task in the literature. However, the mention-queries in those collections, besides being relatively few in number, are limited in that all refer to people for whom a substantial amount of evidence would be expected to be available in the collection thus omitting the "long tail'' of the identity distribution for which less evidence is available. This motivated the development of a new test collection that now is the largest and best-balanced test collection available for the task. To build this collection, a user study was conducted that also provided some insight into the difficulty of the task and how time-consuming it is when humans perform it, and the reliability of their task performance. The study revealed that at least 80% of the 584 annotated mentions were resolvable to people who had sent or received email within the same collection. The new test collection was used to experimentally evaluate the resolution system. The results highlight the importance of the social context (that includes messages sent or received by the original communicating parties) when resolving mentions in email. Moreover, the results show that combining evidence from multiple types of contexts yields better resolution than what can be achieved using any individual context. The one-best selection is correct 74% of the time when tested on the full set of the mention-queries, and 51% of the time when tested on the mention-queries labeled as "hard'' by the annotators. Experiments run with iterative reformulation of the resolution algorithm resulted in modest gains only for the second iteration in the social context expansion.