Learning Term-weighting Functions for Similarity Measures

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 793–802, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Learning Term-weighting Functions for Similarity Measures Wen-tau Yih Microsoft Research Redmond, WA, USA scottyih@microsoft.com Abstract Measuring the similarity between two texts is a fundamental problem in many NLP and IR applications. Among the ex- isting approaches, the cosine measure of the term vectors representing the origi- nal texts has been widely used, where the score of each term is often determined by a TFIDF formula. Despite its sim- plicity, the quality of such cosine similar- ity measure is usually domain dependent and decided by the choice of the term- weighting function. In this paper, we pro- pose a novel framework that learns the term-weighting function. Given the la- beled pairs of texts as training data, the learning procedure tunes the model pa- rameters by minimizing the specified loss function of the similarity score. Com- pared to traditional TFIDF term-weighting schemes, our approach shows a significant improvement on tasks such as judging the quality of query suggestions and filtering irrelevant ads for online advertising. 1 Introduction Measuring the semantic similarity between two texts is an important problem that has many use- ful applications in both NLP and IR communi- ties. For example, Lin (1998) defined a similar- ity measure for automatic thesaurus creation from a corpus. Mihalcea et al. (2006) developed sev- eral corpus-based and knowledge-based word sim- ilarity measures and applied them to a paraphrase recognition task. In the domain of web search, dif- ferent methods of measuring similarity between short text segments have recently been proposed for solving problems like query suggestion and al- ternation (Jones et al., 2006; Sahami and Heilman, 2006; Metzler et al., 2007; Yih and Meek, 2007). Among these similarity measures proposed in various applications, the vector-based methods are arguably the most widely used. In this approach, the text being compared with is first represented by a term vector, where each term is associated with a weight that indicates its importance. The similarity function could be cosine (i.e., the inner product of two normalized unit term vectors, or equivalently a linear kernel), or other kernel func- tions such as the Gaussian kernel. There are essentially two main factors that de- cide the quality of a vector-based similarity mea- sure. One is the vector operation that takes as in- put the term vectors and computes the final simi- larity score (e.g., cosine). The other is how these term vectors are constructed, including the term selection process and how the weights are deter- mined. For instance, a TFIDF scheme for mea- suring document similarity may follow the bag-of- words strategy to include all the words in the doc- ument when constructing the term vectors. The weight of each term is simply the product of its term frequency (i.e., the number of occurrences in the document) and inverse document frequency (i.e., the number of documents in a collection that contain this term). Despite its simplicity and …

Content truncated to 500 words