A Beam-Search Decoder for Grammatical Error Correction

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 568–578, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics A Beam-Search Decoder for Grammatical Error Correction Daniel Dahlmeier1 and Hwee Tou Ng1,2 1NUS Graduate School for Integrative Sciences and Engineering 2Department of Computer Science, National University of Singapore {danielhe,nght}@comp.nus.edu.sg Abstract We present a novel beam-search decoder for grammatical error correction. The decoder iteratively generates new hypothesis correc- tions from current hypotheses and scores them based on features of grammatical correctness and fluency. These features include scores from discriminative classifiers for specific er- ror categories, such as articles and preposi- tions. Unlike all previous approaches, our method is able to perform correction of whole sentences with multiple and interacting er- rors while still taking advantage of powerful existing classifier approaches. Our decoder achieves an F1 correction score significantly higher than all previous published scores on the Helping Our Own (HOO) shared task data set. 1 Introduction Grammatical error correction is an important prob- lem in natural language processing (NLP) that has attracted an increasing amount of interest over the last few years. Grammatical error correction promises to provide instantaneous accurate feedback to language learners, e.g., learners of English as a Second Language (ESL). The dominant paradigm that underlies most er- ror correction systems to date is multi-class clas- sification. A classifier is trained to predict a word from a confusion set of possible correction choices, given some feature representation of the surround- ing sentence context. During testing, the classifier predicts the most likely correction for each test in- stance. If the prediction differs from the observed word used by the writer and the classifier is suffi- ciently confident in its prediction, the observed word is replaced by the prediction. Although considerable progress has been made, the classification approach suffers from some serious shortcomings. Each clas- sifier corrects a single word for a specific error cat- egory individually. This ignores dependencies be- tween the words in a sentence. Also, by conditioning on the surrounding context, the classifier implicitly assumes that the surrounding context is free of gram- matical errors, which is often not the case. Finally, the classifier typically has to commit to a single one- best prediction and is not able to change its deci- sion later or explore multiple corrections. Instead of correcting each word individually, we would like to perform global inference over corrections of whole sentences which can contain multiple and interact- ing errors. An alternative paradigm is to view error correc- tion as a statistical machine translation (SMT) prob- lem from “bad” to “good” English. While this ap- proach can naturally correct whole sentences, a stan- dard SMT system cannot easily incorporate mod- els for specific grammatical errors. It also suffers from the paucity of error-annotated training data for grammar correction. As a result, applying a stan- dard SMT system to error correction does not pro- duce good results, as we show in this work. …

Content truncated to 500 words