Query Expansion for IR using Knowledge-Based Relatedness

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1467–1471, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Query Expansion for IR using Knowledge-Based Relatedness Arantxa Otegi IXA NLP Group Univ. of the Basque Country arantza.otegi@ehu.es Xabier Arregi IXA NLP Group Univ. of the Basque Country xabier.arregi@ehu.es Eneko Agirre IXA NLP Group Univ. of the Basque Country e.agirre@ehu.es Abstract The limitations of keyword-only ap- proaches to information retrieval were rec- ognized since the early days, specially in cases where different but closely-related words are used in the query and the rel- evant document. Query expansion tech- niques like pseudo-relevance feedback rely on the target document set in order to bridge the gap between those words, but they might suffer from topic drift. This pa- per explores the use of knowledge-based semantic relatedness in order to bridge the gap between query and documents. We performed query expansion, with pos- itive effects over some language modeling baselines. 1 Introduction The potential pitfalls of keyword retrieval have been noted since the earliest days of Information Retrieval (IR). Keyword retrieval proves ineffec- tive when different but closely-related words are used in the query and the relevant document. The use of different words creates a lexical gap be- tween the query and the document. In order to bridge the gap, IR has resorted to distributional semantic models. Most research concentrated on Query Expansion (QE) methods, which typically analyze term co-occurrence statis- tics in the corpus and/or in the highest scored doc- uments in order to select terms for expanding the query terms (Manning et al., 2009). The work pre- sented here is complementary, in that we explore QE, but we use an approach based on semantic re- latedness instead of distributional methods. In a closely related work, (Agirre et al., 2010) proposed a WordNet-based document expansion method using random walks: given a document, a random walk algorithm over the WordNet graph, inspired in (Agirre et al., 2009b), ranks concepts closely related to the words in the document. Note that the method can return concepts which are not explicitly mentioned in the document. The highest ranking concepts were then selected to expand the document. In this work, we explore an alternative method to exploit relatedness, query expansion, so we thus run the relatedness algorithm over the queries and we expand the queries. We adopt a language modeling framework to implement the query like- lihood and pseudo-relevance feedback baselines, as well as our relatedness-based query expansion method. In order to test the performance of our method we selected several datasets with different do- mains, topic typologies and document lengths. Given the relevance among the community us- ing WordNet-related methods, we selected the Robust-WSD dataset from CLEF (Agirre et al., 2009a), which is a typical ad-hoc dataset on news. As we think that our method is specially relevant for short queries and/or short documents, we also selected the Yahoo! Answers dataset, which con- tains questions and answers as phrased by real users on …

Content truncated to 500 words