Experiments with Language Models for Known-Item Finding of E-mail Messages

We present experiments using language models to rank e-mail messages for the Known-Item Finding task of the Enterprise track. We combine evidence from the text of the message, its subject, the text of the thread the in which the message occurs, and the text of messages that are in reply to the message. We find that the only statistically significant differences suggest that in addition to the text of the message, the subject is a very important piece of evidence. We also explore the use of a depth based prior, where emphasis is place on messages near the root of the thread structure, which has mixed results.