Statistical Source Expansion for Question Answering

A source expansion algorithm automatically extends a given text corpus with related information from large, unstructured sources. While the expanded corpus is not intended for human consumption, it can be leveraged in question answering (QA) and other information retrieval or extraction tasks to find more relevant knowledge and to gather additional evidence for evaluating hypotheses. In this thesis, we propose a novel algorithm that expands a collection of seed documents by (1) retrieving related content from the Web or other large external sources, (2) extracting self-contained text nuggets from the related content, (3) estimating the relevance of the text nuggets with regard to the topics of the seed documents using a statistical model, and (4) compiling new pseudo-documents from nuggets that are relevant and complement existing information. In an intrinsic evaluation on a dataset comprising 1,500 hand-labeled web pages, the most elective statistical relevance model ranked text nuggets by relevance with 81% MAP, compared to 43% when relying on rankings generated by a web search engine, and 75% when using a multi-document summarization algorithm. These differences are statistically significant and result in noticeable gains in search performance in a task-based evaluation on QA datasets. The statistical models use a comprehensive set of features to predict the topicality and quality of text nuggets based on topic models built from seed content, search engine rankings and surface characteristics of the retrieved text. Linear models that evaluate text nuggets individually are compared to a sequential model that estimates their relevance given the surrounding nuggets. The sequential model leverages features derived from text segmentation algorithms to dynamically predict transitions between relevant and irrelevant passages. It slightly outperforms the best linear model while using fewer parameters and requiring less training time. In addition, we demonstrate that active learning reduces the amount of labeled data required to fit a relevance model by two orders of magnitude with little loss in ranking performance. This facilitates the adaptation of the source expansion algorithm to new knowledge domains and applications. Applied to the QA task, the proposed method yields consistent and statistically significant performance gains across different datasets, seed corpora and retrieval strategies. We evaluated the impact of source expansion on search performance and end-to-end accuracy using Watson and the OpenEphyra QA system, and datasets comprising over 6,500 questions from the Jeopardy! quiz show and TREC evaluations. By expanding various seed corpora with web search results, we were able to improve the QA accuracy of Watson from 66% to 71% on regular Jeopardy! questions, from 45% to 51% on Final Jeopardy! questions and from 59% to 64% on TREC factoid questions. We also show that the source expansion approach can be adapted to extract relevant content from locally stored sources without requiring a search engine, and that this method yields similar performance gains. When combined with the approach that uses web search results, Watson's accuracy further increases to 72% on regular Jeopardy! data, 54% on Final Jeopardy! and 67% on TREC questions.