Faculty Recruiting Make a Gift

Topic Regression

13 Sep
Tuesday, 09/13/2011 6:00am to 8:00am
Ph.D. Thesis Defense

David Mimno

Computer Science Building, Room 151

Text documents are generally accompanied by non-textual information, such as authors, dates, publication sources, and, increasingly, automatically recognized named entities. Work in text analysis has often involved predicting these non-text values based on text data for tasks such as document classification and author identification. This thesis considers the opposite problem: predicting the textual content of documents based on non-text data. In this work I study several regression-based methods for estimating the influence of specific metadata elements in determining the content of text documents. Such topic regression methods allow users of document collections to test hypotheses about the underlying environments that produced those documents.

Advisor: Andrew McCallum