Grammatical Error Correction Using Feature Selection and Confidence Tuning

International Joint Conference on Natural Language Processing, pages 1067–1071, Nagoya, Japan, 14-18 October 2013. Grammatical Error Correction Using Feature Selection and Confidence Tuning Yang Xiang1†, Yaoyun Zhang1, Xiaolong Wang1‡, Chongqiang Wei1, Wen Zheng1, Xiaoqiang Zhou1, Yuxiu Hu2, Yang Qin1 1Key Laboratory of Network Oriented Intelligent Computation, Harbin Institute of Technology Shenzhen Graduate School, China 2South University of Science and Technology of China †windseed@gmail.com ‡wangxl@insun.hit.edu.cn Abstract This paper proposes a novel approach to re- solve the English article error correction prob- lem, which accounts for a large proportion in grammatical errors. Most previous machine learning based researches empirically collect- ed features which may bring about noises and increase the computational complexity. Meanwhile, the predicted result is largely af- fected by the threshold setting of a classifier which can easily lead to low performance but hasn’t been well developed yet. To address these problems, we employ genetic algorithm for feature selection and confidence tuning to reinforce the motivation of correction. Com- parative experiments on the NUCLE corpus show that our approach could efficiently re- duce feature dimensionality and enhance the final F1 value for the article error correction problem. 1 Introduction Grammatical errors in English are common in written issues especially for learners of English as a second language (L2 leaners). As a result, automatic grammatical error correction (GEC) sprung up and has attracted more and more re- search attention recently. Among various error types, article errors account for a large propor- tion (over 12% in NUCLE) and are very difficult to be corrected. Articles in the English language include indef- inite article a and an, definite article the and zero article empty which means no article is used in this position. Articles are determiners of noun phrases which are indispensable in English grammar. Article errors are common in written English including wrong use, missing, and un- necessary use of articles. For example, in the fol- lowing sentence“Over these years, it had helped humans to improve the accessibility in the forms of cards to gain access to certain places.”there are two thes in which the first one is required but the second one is unnecessary. It is difficult for L2 learners to judge whether an article is neces- sary or not, or which article is needed. These er- rors are highly correlated with the context fea- tures around noun phrases. Errors occur fre- quently in various written issues which motivates researchers to exploit automatic error correction. There are two main approaches for English ar- ticle error correction. One of them is the external language materials based approach. Although there are minor differences on strategies, the main idea of this approach is to use frequencies such as n-gram counts as a filter and keep those phrases that have relatively high frequencies. Typical researches are shown by (Yi et al., 2008) and (Bergsma et al., 2009). Similar methods also exist in HOO shared tasks1 such as the web 1TB n-gram features used by (Dahlmeier and Ng, 2012a) and the large-scale n-gram model in (Heilman …

Content truncated to 500 words