Native Language Identification and Reconstruction of Native Language Relationship Using Japanese Learner Corpus

Native Language Identification and Reconstruction of Native Language Relationship Using Japanese Learner Corpus Mitsuhiro Nishijima1,2 and Ying Liu1 1Department of Chinese Language and Literature, Tsinghua University, China 2Department of Industrial Engineering and Economics, Tokyo Institute of Technology, Japan nishijima.m.ae@m.titech.ac.jp, yingliu@tsinghua.edu.cn Abstract We investigated the influence of the native language (L1) and its relationship preserved in second language (L2) Japanese texts by machine learning methods. We firstly per- formed native language identification (NLI) on L2 Japanese texts and proposed the char- acter type as a new Japanese-specific feature in NLI. Combining it with other features in- creases the identification accuracy, and we ob- tained the maximum accuracy of 72.2%. We also confirmed the L1 relationship from the confusion matrix of NLI. To investigate the L1 relationship in more detail, we secondly ap- plied hierarchical clustering to Japanese texts. Based on the results, we proposed a new hy- pothesis about the L1 relationship preserved in L2 texts. 1 Introduction When we write a text in a second language (L2), the native language (L1) has more or less positive or negative impact on the L2 text, and traces of the L1 remain on the text. Therefore, many studies have been conducted to identify the L1 of the writer from L2 texts using machine learning methods—this is native language identification (NLI). NLI can be applied to automatic error correction on texts written by L2 learners. It has been pointed out that the quality of the correction can be improved by taking the writers’ L1 into account (Rozovskaya and Roth, 2011). If the writers’ L1 is automatically identified by NLI, then the correction system can provide better quality feedback to the L2 learners. NLI can also be regarded as a kind of author pro- filing, and it is expected to use for criminal inves- tigations and marketing (Estival et al., 2007). The number of L2 Japanese learners is increasing every year (Japan Foundation, 2018). Therefore, it is im- portant to extend NLI, which has been conducted mainly on L2 English, to L2 Japanese. In general, it can be observed from the identifica- tion results of NLI that L1s sharing a geographical or genealogical relationship are easily misidentified. It means that L2 texts preserve information not only on L1 itself but also on the L1 relationship. Some studies have successfully reconstructed the L1 rela- tionship from L2 English texts (see Section 2 in de- tail). However, all of them used L2 English texts, and thus whether the preservation of the L1 relation- ship on L2 texts depends on the L2 has not been dis- cussed yet. Although Nagata (2014) discussed the dependence on L1, the argumentation was incom- plete. Based on the above problems, we firstly perform NLI on L2 Japanese texts in Section 3. We propose the character type as a Japanese-specific feature in NLI. We confirm that combining it with other fea- tures improves the identification accuracy. The max- imum accuracy reaches 72.2% against a baseline of 8.3%. We also confirm the L1 relationship …

Content truncated to 500 words