Sentential Paraphrase Generation for Agglutinative Languages Using SVM with a String Kernel

PACLIC 28 !650 Sentential Paraphrase Generation for Agglutinative Languages Using SVM with a String Kernel Hancheol Park1, Gahgene Gweon1, Ho-Jin Choi2, Jeong Heo3, Pum-Mo Ryu3 1Department of Knowledge Service Engineering, KAIST 2Department of Computer Science, KAIST 3Knowledge Mining Research Team, ETRI {hancheol.park, ggweon, hojinc}@kaist.ac.kr {jeonghur, pmryu}@etri.re.kr Abstract Paraphrase generation is widely used for various natural language processing (NLP) applications such as question answering, multi-document summarization, and machine translation. In this study, we identify the problems occurring in the process of applying existing probabilistic model-based methods to agglutinative languages, and provide solutions by reflecting the inherent characteristics of agglutinative languages. More specifically, we propose and evaluate a sentential paraphrase generation (SPG) method for the Korean language using Support Vector Machines (SVM) with a string kernel. The quality of generated paraphrases is evaluated using three criteria: (1) meaning preservation, (2) grammaticality, and (3) equivalence. Our experiment shows that the proposed method outperformed a probabilistic model-based method by 12%, 16%, and 17%, respectively, with respect to the three criteria. 1 Introduction Paraphrase generation (PG) is a useful technique in various natural language processing (NLP) applications, where it expands natural language expressions. In question answering systems, PG can be utilized to generate semantically equivalent questions. It can solve word mismatch problems when searching for answers (Lin and Pantel, 2001; Riezler et al., 2007). For multi-document summarization, it also helps to generate a summary sentence by identifying repeated information among semantically similar sentences (McKeown et al., 2002). In addition, for machine translation, paraphrasing can mitigate the scarcity of training data by expanding the reference translations (Callison-Burch, 2006). In this study, we focus on a paraphrase generation approach, namely, sentential paraphrase generation (SPG), which takes a whole sentence as an input and generates a paraphrased output sentence that has the same meaning. Figure 1 shows an overview of the SPG process in general. Figure 1: Example of the SPG process. For example, let us assume that we would like to generate a paraphrased sentence using bilingual parallel corpora for a Korean input sentence “삼성본사는 서울에 위치하고 있다 (The headquarters of Samsung is located in Seoul).” For simplicity, in the examples used in this paper, we assume that an input sentence has only one source phrase to be substituted/ paraphrased. In our sample sentence, the source phrase is “위치하고 있다 (is located).” Currently, popular methods for SPG use phrase-based statistical machine Copyright 2014 by Hancheol Park, Gahgene Gweon, Ho-Jin Choi, Jeong Heo, and Pum-Mo Ryu 28th Pacific Asia Conference on Language, Information and Computation pages 650–657 PACLIC 28 !651 translation (PBSMT) techniques (Bannard and Callison-Burch, 2005; Callison-Burch, 2008; Zhao et al., 2009; Wubben et al., 2010) with phrase- based paraphrase sets extracted from bilingual or monolingual parallel corpora. Such methods based on PBSMT use probabilistic-based models (e.g., a paraphrase model (PM) and a language model (LM)) to select the best phrase for substitution from a paraphrase set, which contains phrases that share the same meaning, to produce a paraphrased sentence. Probabilistic-based methods improve the system as the …

Content truncated to 500 words