Mining Revision Log of Language Learning SNS for Automated Japanese Error Correction of Second Language Learners

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 147–155, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Mining Revision Log of Language Learning SNS for Automated Japanese Error Correction of Second Language Learners Tomoya Mizumoto NAIST, Japan tomoya-m@is.nasit.jp Mamoru Komachi NAIST, Japan komachi@is.naist.jp Masaaki Nagata NTT, Japan nagata.masaaki@lab.ntt.co.jp Yuji Matsumoto NAIST, Japan matsu@is.naist.jp Abstract We present an attempt to extract a large- scale Japanese learners’ corpus from the revision log of a language learning SNS. This corpus is easy to obtain in large- scale, covers a wide variety of topics and styles, and can be a great source of knowl- edge for both language learners and in- structors. We also demonstrate that the extracted learners’ corpus of Japanese as a second language can be used as train- ing data for learners’ error correction us- ing an SMT approach. We evaluate dif- ferent granularities of tokenization to al- leviate the problem of word segmentation errors caused by erroneous input from lan- guage learners. Experimental results show that the character-wise model outperforms the word-wise model. 1 Introduction The number of Japanese language learners around the world has increased more than 30-fold in the past three decades. The Japan Foundation reports that more than 3.65 million people in 133 coun- tries and regions are studying Japanese in 2009 1. However, there are only 50,000 Japanese language teachers overseas, and thus it is in high demand to find good instructors for writers of Japanese as a Second Language (JSL). Recently, natural language processing research has begun to pay attention to second language learning (Rozovskaya and Roth, 2011; Park and Levy, 2011; Liu et al., 2011; Oyama and Mat- sumoto, 2010; Xue and Hwa, 2010). However, most previous research for second language learn- ing deals with restricted types of learners’ er- rors. For example, research for JSL learners’ 1http://www.jpf.go.jp/e/japanese/ survey/result/index.html errors mainly focus on Japanese case particles (Oyama and Matsumoto, 2010; Imaeda et al., 2003; Nampo et al., 2007; Suzuki and Toutanova, 2006), however they focus only on case particles whereas we attempt to correct all types of errors. However, real JSL learners’ writing contains not only errors of Japanese case particles but also various other errors including spelling and collo- cation errors. For instance, a Japanese language learner who speaks Chinese may write: 何で日本語はこんなに難しい な の？ (Why does Japanese are so difficult?) which has a grammatical error of inserting ‘な’ due to literal translation from Chinese. Park and Levy (2011) proposed an EM-based unsuper- vised approach to perform whole sentence gram- mar correction, but the types of errors must be pre- determined to learn the parameters for their noisy channel model. It requires expert knowledge of L2 teaching, which is often hard to obtain. One promising approach for correcting unre- stricted errors of JSL learners is Brockett et al. (2006)’s automated error correction method using statistical machine translation (SMT). The advan- tage of their method is that it does not require expert knowledge. Instead, it learns a correction …

Content truncated to 500 words