Vocabulary Filtering for Term Weighting in Archived Question Search

摘要

This paper proposes the notion of vocabulary filtering in a term weighting framework that consists of three filters at the document level, collection level, and vocabulary level. While term frequency and document frequency along with their variations are respectively the dominant term weighting factors at the document level and collection level, vocabulary level factors are seldom considered in current models. In a way, stopword removal can be seen as a vocabulary level filter, but it is not well integrated into the current term-weighting models. In this paper, we propose a vocabulary filtering and multi-level term weighting model by integrating point-wise divergence based measure into the commonly used TF-IDF model. With our proposed model, the specificity of the vocabulary is captured as a new factor in term weighting, and stopwords are naturally handled within the model rather than being removed according to a separately constructed list. Experiments conducted on searching for similar questions in a large community-based question answering archive show that: (a)our proposed term weighting model with multiple levels is consistently better than those with single level for retrieval task; (b)the proposed vocabulary filter well distinguishes salient and trivial terms, and can be utilized to construct stopword lists.