IR Series - Kevyn Collins-Thompson - Friday, October 5 - 12:00 (noon), NSH 3002

[Jaime Arguello]

Greetings,

Please join us for our first IR Series talk this fall!

Lunch will be provided by Yahoo!

Time and Location:
Friday, October 5, 12:00 - 1:00 pm, Newell-Simon Hall (NSH) 3002

Speaker:
Kevyn Collins-Thompson

Title:
Estimating and Exploiting Uncertainty in Pseudo-Relevance Feedback

Abstract:
We give an overview of our recent research on methods for
quantifying uncertainty in retrieval models and exploiting this new
information for improved retrieval performance.  We focus on the case of
pseudo-relevance feedback, which is an automatic method for enhancing an
initial query with additional terms by using the top-retrieved documents
as a type of noisy training set for relevance.  Current feedback
algorithms typically improve precision on average, but can be very
unstable for individual queries, with poor worst-case performance.

We show how sampling methods can be used to improve both the robustness
(worst-case performance) and precision of a strong baseline feedback
algorithm, and provide principled sensitivity estimates for arbitrary
feedback functions.  We demonstrate how these estimates can in turn be
used in applications such as predicting query difficulty or forming useful
risk constraints in query model optimization problems.

This talk includes joint work with Jamie Callan.