Go Climb a Dependency Tree and Correct the Grammatical Errors

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 266–277, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Go Climb a Dependency Tree and Correct the Grammatical Errors Longkai Zhang Houfeng Wang Key Laboratory of Computational Linguistics (Peking University) Ministry of Education, China zhlongk@qq.com, wanghf@pku.edu.cn Abstract State-of-art systems for grammar error correction often correct errors based on word sequences or phrases. In this paper, we describe a grammar error correction system which corrects grammatical errors at tree level directly. We cluster all error into two groups and divide our system into two modules correspondingly: the general module and the special module. In the general module, we propose a TreeNode Language Model to correct errors related to verbs and nouns. The TreeNode Lan- guage Model is easy to train and the de- coding is efficient. In the special module, two extra classification models are trained to correct errors related to determiners and prepositions. Experiments show that our system outperforms the state-of-art sys- tems and improves the F1 score. 1 Introduction The task of grammar error correction is difficult yet important. An automatic grammar error cor- rection system can help second language (L2) learners improve the quality of their writing. In re- cent years, there are various competitions devoted to grammar error correction, such as the HOO- 2011(Dale and Kilgarriff, 2011), HOO-2012(Dale et al., 2012) and the CoNLL-2013 shared task (Ng et al., 2013). There has been a lot of work ad- dressing errors made by L2 learners. A significant proportion of the systems for grammar error cor- rection train individual statistical models to cor- rect each special kind of error word by word and ignore error interactions. These methods assume no interactions between different kinds of gram- matical errors. In real problem settings errors are correlated, which makes grammar error correction much more difficult. Recent research begins to focus on the error interaction problem. For example, Wu and Ng (2013) decodes a global optimized result based on the individual correction confidence of each kind of errors. The individual correction confi- dence is still based on the noisy context. Ro- zovskaya and Roth (2013) uses a joint modeling approach, which considers corrections in phrase structures instead of words. For dependencies that are not covered by the joint learning model, Ro- zovskaya and Roth (2013) uses the results of Illi- nois system in the joint inference. These results are still at word level and are based on the noisy context. These systems can consider error inter- actions, however, the systems are complex and inefficient. In both Wu and Ng (2013) and Ro- zovskaya and Roth (2013), Integer Linear Pro- gramming (ILP) is used for decoding a global op- timized result. In the worst case, the time com- plexity of ILP can be exponent. In contrast, we think a better grammar error cor- rection system should correct grammatical errors at sentence level directly and efficiently. The sys- tem should correct as many kinds of errors as pos- …

Content truncated to 500 words