Similarity measures and diversity rankings for query-focused sentence extraction

Palakorn Achananuparp

doi:10.17918/etd-3245

Query-focused sentence extraction generally refers to an extractive approach to select a set of sentences that responds to a specific information need. It is one of the major approaches employed in multi-document summarization, focused summarization, and complex question answering. The major advantage of most extractive methods over the natural language processing (NLP) intensive methods is that they are relatively simple, theoretically sound - drawing upon several supervised and unsupervised learning techniques, and often produce equally strong empirical performance. Many research areas, including information retrieval and text mining, have recently moved toward the extractive query-focused sentence generation as its outputs have great potential to support every day's information seeking activities. Particularly, as more information have been created and stored online, extractive-based summarization systems may quickly utilize several ubiquitous resources, such as Google search results and social medias, to extract summaries to answer users' queries. This thesis explores how the performance of sentence extraction tasks can be improved to create higher quality outputs. Specifically, two major areas are investigated. First, we examine the issue of natural language variation which affects the similarity judgment of sentences. As sentences are much shorter than documents, they generally contain fewer occurring words. Moreover, the similarity notions of sentences are different than those of documents as they tend to be very specific in meanings. Thus many document-level similarity measures are likely to perform well at this level. In this work, we address these issues in two application domains. First, we present a hybrid method, utilizing both unsupervised and supervised techniques, to compute the similarity of interrogative sentences for factoid question reuse. Next, we propose a novel structural similarity measure based on sentence semantics for paraphrase identification and textual entailment recognition tasks. The empirical evaluations suggest the effectiveness of the proposed methods in improving the accuracy of sentence similarity judgments. Furthermore, we examine the effects of the proposed similarity measure in two specific sentence extraction tasks, focused summarization and complex question answering. In conjunction with the proposed similarity measure, we also explore the issues of novelty, redundancy, and diversity in sentence extraction. To that end, we present a novel approach to promote diversity of extracted sets of sentences based on the negative endorsement principle. Negative-signed edges are employed to represent a redundancy relation between sentence nodes in graphs. Then, sentences are reranked according to the long-term negative endorsements from random walk. Additionally, we propose a unified centrality ranking and diversity ranking based on the aforementioned principle. The results from a comprehensive evaluation confirm that the proposed methods perform competitively, compared to many state-of-the-art methods.

Similarity measures and diversity rankings for query-focused sentence extraction

Files and links (1)

Abstract

Metrics

Details

Similarity measures and diversity rankings for query-focused sentence extraction

Files and links (1)

Abstract

Metrics

Details

Drexel University Social media