Building Queryable Datasets from Ungrammatical and Unstructured Sources

For agents to act on behalf of users, they will have to query the vast amounts of textual data on the internet. However, much of this text cannot be queried because it is neither grammatical nor formally structured enough to support traditional information extraction approaches to annotation. Examples of such text, called “posts,” include item descriptions on Ebay or internet classifieds like Craig’s list. This work describes an approach to annotating posts by combining record linkage with information extraction. This approach leverages collections of known entities, called “reference sets,” by first aligning a post to a member of a reference set, and then exploiting this matched member during information extraction. This thesis compares this extraction approach to more traditional information extraction methods that rely on structural and grammatical characteristics, and it shows that this approach outperforms traditional methods on this type of data.