Chinese Preposition Selection for Grammatical Error Diagnosis

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 888–899, Osaka, Japan, December 11-17 2016. Chinese Preposition Selection for Grammatical Error Diagnosis Hen-Hsen Huang, Yen-Chi Shao, and Hsin-Hsi Chen Department of Computer Science and Information Engineering National Taiwan University No. 1, Sec. 4, Roosevelt Road, Taipei, 10617 Taiwan {hhhuang, ycshao}@nlg.csie.ntu.edu.tw; hhchen@ntu.edu.tw Abstract Misuse of Chinese prepositions is one of common word usage errors in grammatical error diag- nosis. In this paper, we adopt the Chinese Gigaword corpus and HSK corpus as L1 and L2 cor- pora, respectively. We explore gated recurrent neural network model (GRU), and an ensemble of GRU model and maximum entropy language model (GRU-ME) to select the best preposition from 43 candidates for each test sentence. The experimental results show the advantage of the GRU models over simple RNN and n-gram models. We further analyze the effectiveness of linguistic information such as word boundary and part-of-speech tag in this task. 1 Introduction As learning Chinese has been popular world-wide, Chinese word spelling checking and grammatical error diagnosis for learners of Chinese language is recently advanced in the NLP community. In the NLP-TEA shared tasks (Yu et al., 2014; Lee et al., 2015a, Lee et al., 2015b), four types of grammatical errors including disorder, redundant, missing, and mis-selection are defined. These tasks focus on de- tecting and identifying grammatical errors, but not addressing error correction. In English, selection of appropriate prepositions is a barrier to non-native language learners (Cho- dorow et al., 2007; Felica and Pulman, 2008). Many studies deal with the issue of preposition error detection and correction (Dale et al., 2012; Ng et al., 2013). The selection of Chinese prepositions is also challenging to non-native learners, especially for some common prepositions. (S1) and (S2) show a real case from the HSK corpus. The preposition 從 (cóng, “from”) in (S2) is preferred to 在 (zài, “on/in/at”) in (S1). In this paper, we investigate Chinese preposition selection. (S1) 在 公眾 利益 方面 來看 (on the view point of public interest) (S2) 從 公眾 利益 方面 來看 (from the view point of public inter- est) A model trained on an error-annotated dataset benefits from learning the mapping between wrong instances and their corrected counterparts (Cahill et al., 2013). However, large error-annotated dataset for Chinese preposition error correction is still not available. In this paper, we utilize a large-scale corpus written by native Chinese speakers to deal with this problem. We adopt language models based on re- current neural networks (RNNs) (Mikolov et al., 2011) to capture word dependencies in sentences. Cut- ting-edge methods like noise contrastive estimation (NCE) (Chen et al., 2015a) and gated recurrent unit (GRU) (Cho et al., 2014) are explored in RNNs. For the task of 43-way classification, i.e., to select the best preposition from 43 candidates, the gated recurrent neural network model (GRU) achieves an ac- curacy of 74.05% and an MRR of 83.08% on the Chinese Gigaword corpus (L1 corpus), and achieves an accuracy of 60.13% and an MRR …

Content truncated to 500 words