PolyUCOMP-CORE_TYPED: Computing Semantic Textual Similarity using Overlapped Senses

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 90–95, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics PolyUCOMP-CORE TYPED: Computing Semantic Textual Similarity using Overlapped Senses Jian Xu Qin Lu The Hong Kong Polytechnic University Department of Computing Hung Hom, Kowloon, Hong Kong {csjxu, csluqin}@comp.polyu.edu.hk Abstract The Semantic Textual Similarity (STS) task aims to exam the degree of semantic equivalence between sentences (Agirre et al., 2012). This paper presents the work of the Hong Kong Polytechnic University (PolyUCOMP) team which has participated in the STS core and typed tasks of SemEval- 2013. For the STS core task, the PolyUCOMP system disambiguates words senses using contexts and then determine sentence similarity by counting the number of senses they shared. For the STS typed task, the string kernel (Lodhi et al., 2002) is used to compute similarity between two entities to avoid string variations in entities. 1 Introduction Sentence similarity computation plays an important role in text summarization and social network applications (Erkan et al., 2004; Jin et al., 2011). The SemEval 2012 competition initiated a task targeted at Semantic Textual Similarity (STS) between sentence pairs (Agirre et al., 2012). Given a set of sentence pairs, participants are required to assign to each sentence pair a similarity score. Because a sentence has only a limited amount of content words, it is difficult to determine sentence similarities. To solve this problem, Hatzivassiloglou et al. (1999) proposed to use linguistic features as indicators of text similarity to address the problem of sparse representation of sentences. Mihalcea et al. (2006) measured sentence similarity using component words in sentences. Li et al. (2006) proposed to incorporate the semantic vector and word order to calculate sentence similarity. Biemann et al. (2012) applied the log-linear regression model by combining the simple string based measures, for example, word ngrams and semantic similarity measures, for example, textual entailment. Similarly, Saric et al. (2012) used a support vector regression model which incorporates features computed from sentence pairs. The features are knowledge- and corpus-based word similarity, ngram overlaps, WordNet augmented word overlap, syntactic features and so on. Xu et al. (2012) combined semantic vectors with skip bigrams to determine sentence similarity, whereas the skip bigrams take into the sequential order between words. In our approach to the STS task, words in sentences are assigned with appropriate senses using their contexts. Sentence similarity is computed by calculating the number of shared senses in both sentences since it is reasonable to assume that similar sentences should have more overlapping senses. For the STS-TYPED task, variations might occur in author names, people involved, time expression and location. Thus, string kernel is applied to compute similarity between entities because it can capture variations between entities. Moreover, for the event similarity in STS-TYPED task, semantic relatedness between verbs is derived the WordNet. The rest of this paper is structured as follows. Section 2 describes sentence similarity using sense overlapping and string kernel. …

Content truncated to 500 words