Discriminative Reranking for Grammatical Error Correction with Statistical Machine Translation

Proceedings of NAACL-HLT 2016, pages 1133–1138, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Discriminative Reranking for Grammatical Error Correction with Statistical Machine Translation Tomoya Mizumoto Tohoku University tomoya-m@ecei.tohoku.ac.jp Yuji Matsumoto Nara Institute of Science and Technology matsu@is.naist.jp Abstract Research on grammatical error correction has received considerable attention. For dealing with all types of errors, grammatical error cor- rection methods that employ statistical ma- chine translation (SMT) have been proposed in recent years. An SMT system generates candidates with scores for all candidates and selects the sentence with the highest score as the correction result. However, the 1-best re- sult of an SMT system is not always the best result. Thus, we propose a reranking approach for grammatical error correction. The rerank- ing approach is used to re-score N-best results of the SMT and reorder the results. Our ex- periments show that our reranking system us- ing parts of speech and syntactic features im- proves performance and achieves state-of-the- art quality, with an F0.5 score of 40.0. 1 Introduction Research on assisting second language learners has received considerable attention, especially regarding grammatical error correction of essays written by English as a Second Language (ESL) learners. To address all types of errors, grammatical error correc- tion methods that use statistical machine translation (SMT) have been proposed (Brockett et al., 2006; Mizumoto et al., 2012; Buys and van der Merwe, 2013; Yuan and Felice, 2013; Felice et al., 2014; Junczys-Dowmunt and Grundkiewicz, 2014). SMT- based error correction systems have achieved rank- ings first and third in the CoNLL2014 Shared Task (Ng et al., 2014). Figure 1: Flow of reranking. SMT systems generate many candidates of trans- lation. SMT systems generate scored candidates and select a sentence having the highest score as the translation result. However, the 1-best result of SMT system is not always the best result because the scor- ing is conducted only with local features. In other words, N-best (N > 1) results may be better than the 1-best result. Reranking approaches have been devised to solve the scoring problem. Reranking is a method that re-scores N-best candidates of SMT and reorders the candidates by score. Figure 1 shows a flow of reranking. First, N-best results are obtained by a grammatical error correction system using SMT for a learner sentence (A in Figure 1). A reranking sys- 1133 tem then re-scores the N-best results and reorders them (B in Figure 1). In this study, we apply a discriminative rerank- ing method to the task of grammatical error cor- rection. Syntactic information is not considered in the phrase-based SMT. We show that using syntac- tic features in the reranking system can improve er- ror correction performance. Although reranking us- ing only surface features (Shen et al., 2004) is not effective for grammatical error correction, reranking using syntactic features improves the F0.5 score. 2 Related Work for Reranking Reranking approaches have been proposed for com- mon SMT tasks (Shen et al., 2004; Carter and Monz, 2011; Li and Khudanpur, …

Content truncated to 500 words