Utilizing sub-topical structure of documents for information retrieval

Text segmentation in natural language processing typically refers to the process of decomposing a document into constituent subtopics. Our work centers on the application of text segmentation techniques within information retrieval (IR) tasks. For example, for scoring a document by combining the retrieval scores of its constituent segments, exploiting the proximity of query terms in documents for ad-hoc search, and for question answering (QA), where retrieved passages from multiple documents are aggregated and presented as a single document to a searcher. Feedback in ad-hoc IR task is shown to benefit from the use of extracted sentences instead of terms from the pseudo relevant documents for query expansion. Retrieval effectiveness for patent prior art search task is enhanced by applying text segmentation to the patent queries. Another aspect of our work involves augmenting text segmentation techniques to produce segments which are more readable with less unresolved anaphora. This is particularly useful for QA and snippet generation tasks where the objective is to aggregate relevant and novel information from multiple documents satisfying user information need on one hand, and ensuring that the automatically generated content presented to the user is easily readable without reference to the original source document.