Sentence Clustering via Projection over Term Clusters

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 38–43, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Sentence Clustering via Projection over Term Clusters Lili Kotlerman, Ido Dagan Bar-Ilan University Israel Lili.Kotlerman@biu.ac.il dagan@cs.biu.ac.il Maya Gorodetsky, Ezra Daya NICE Systems Ltd. Israel Maya.Gorodetsky@nice.com Ezra.Daya@nice.com Abstract This paper presents a novel sentence cluster- ing scheme based on projecting sentences over term clusters. The scheme incorporates exter- nal knowledge to overcome lexical variability and small corpus size, and outperforms com- mon sentence clustering methods on two real- life industrial datasets. 1 Introduction Clustering is a popular technique for unsupervised text analysis, often used in industrial settings to ex- plore the content of large amounts of sentences. Yet, as may be seen from the results of our research, widespread clustering techniques, which cluster sen- tences directly, result in rather moderate perfor- mance when applied to short sentences, which are common in informal media. In this paper we present and evaluate a novel sentence clustering scheme based on projecting sentences over term clusters. Section 2 briefly overviews common sentence clustering approaches. Our suggested clustering scheme is presented in Section 3. Section 4 describes an implementation of the scheme for a particular industrial task, followed by evaluation results in Section 5. Section 6 lists directions for future research. 2 Background Sentence clustering aims at grouping sentences with similar meanings into clusters. Commonly, vector similarity measures, such as cosine, are used to de- fine the level of similarity over bag-of-words encod- ing of the sentences. Then, standard clustering algo- rithms can be applied to group sentences into clus- ters (see Steinbach et al. (2000) for an overview). The most common practice is representing the sentences as vectors in term space and applying the K-means clustering algorithm (Shen et al. (2011); Pasquier (2010); Wang et al. (2009); Nomoto and Matsumoto (2001); Boros et al. (2001)). An alterna- tive approach involves partitioning a sentence con- nectivity graph by means of a graph clustering algo- rithm (Erkan and Radev (2004); Zha (2002)). The main challenge for any sentence clustering approach is language variability, where the same meaning can be phrased in various ways. The shorter the sentences are, the less effective becomes exact matching of their terms. Compare the fol- lowing newspaper sentence ”The bank is phasing out the EZ Checking package, with no monthly fee charged for balances over $1,500, and is instead offering cus- tomers its Basic Banking account, which carries a fee” with two tweets regarding the same event: ”Whats wrong.. charging $$ for checking a/c” and ”Now they want a monthly fee!”. Though each of the tweets can be found similar to the long sentence by exact term matching, they do not share any single term. Yet, knowing that the words fee and charge are semanti- cally related would allow discovering the similarity between the two tweets. External resources can be utilized to provide such kind of knowledge, by which sentence representa- tion can be enriched. Traditionally, WordNet (Fell- baum, 1998) has …

Content truncated to 500 words