Sentence Similarity based on Dependency Tree Kernels for Multi-document Summarization

Sentence Similarity based on Dependency Tree Kernels for Multi-document Summarization S¸aziye Bet¨ul ¨Ozates¸∗, Arzucan ¨Ozg¨ur∗, Dragomir R. Radev† ∗Department of Computer Engineering, Bo˘gazic¸i University 34342, Bebek, Istanbul, Turkey {saziye.bilgin, arzucan.ozgur}@boun.edu.tr †Department of EECS, University of Michigan 3917 Beyster Building, Ann Arbor, MI 48109 radev@umich.edu Abstract We introduce an approach based on using the dependency grammar representations of sentences to compute sentence similarity for extractive multi-document summarization. We adapt and investigate the effects of two untyped dependency tree kernels, which have originally been proposed for relation extraction, to the multi-document summarization problem. In addition, we propose a series of novel dependency grammar based kernels to better represent the syntactic and semantic similarities among the sen- tences. The proposed methods incorporate the type information of the dependency relations for sentence similarity calculation. To our knowledge, this is the first study that investigates using dependency tree based sentence similarity for multi-document summarization. Keywords: multi-document summarization, sentence similarity, dependency grammars 1. Introduction Multi-document summarization (MDS), which refers to the task of automatically generating a summary of multiple documents about the same topic without losing the most important information, is one of the most promising solu- tions proposed to overcome the information overload prob- lem (Li et al., 2007). Sentence similarity calculation is a crucial task for many extractive approaches to MDS (Erkan and Radev, 2004; Mihalcea and Tarau, 2005; Wang et al., 2008; Aliguliyev, 2009). Most of them use a bag of words model to com- pute sentence similarity. However, the bag of words model is sometimes inadequate for capturing the syntactic and se- mantic similarities among the sentences, which may affect the qualities of the summaries. To address this problem, we propose to employ dependency grammars, which represent the syntactic dependencies among the words in a sentence, for sentence similarity computation in MDS. Using this ap- proach, concepts in multiple documents and relations be- tween similar contents can be captured. In this study, we first adapt two dependency tree based sen- tence similarity kernels in order to use them in MDS. These kernels are proposed by Culotta and Sorensen (2004) and Choi and Kim (2013) respectively, for relation extraction. They do not take the dependency relation types into account while calculating sentence similarity. We then propose a series of new sentence similarity kernels based on typed dependency grammars and test these kernels on LexRank (Erkan and Radev, 2004), a well-known and publicly avail- able MDS system, by replacing its tf-idf based cosine sim- ilarity method with each of the kernels. The proposed sim- ilarity kernels make use of the binary dependency relations in the sentences. We conduct experiments on DUC 2003 and DUC 2004 Task 2 data sets. The results show that the best one of the proposed methods outperforms the other un- typed tree kernels and LexRank’s own sentence similarity method in terms of ROUGE-1 and ROUGE-2 scores. 2. Related Work Several methods including supervised approaches (Das and Martins, 2007; Pei et al., 2012), topic driven mod- els (Nastase, 2008; Hennig and …

Content truncated to 500 words