Training Paradigms for Correcting Errors in Grammar and Usage

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 154–162, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Training Paradigms for Correcting Errors in Grammar and Usage Alla Rozovskaya and Dan Roth University of Illinois at Urbana-Champaign Urbana, IL 61801 {rozovska,danr}@illinois.edu Abstract This paper proposes a novel approach to the problem of training classifiers to detect and correct grammar and usage errors in text by selectively introducing mistakes into the train- ing data. When training a classifier, we would like the distribution of examples seen in train- ing to be as similar as possible to the one seen in testing. In error correction problems, such as correcting mistakes made by second lan- guage learners, a system is generally trained on correct data, since annotating data for train- ing is expensive. Error generation methods avoid expensive data annotation and create training data that resemble non-native data with errors. We apply error generation methods and train classifiers for detecting and correcting arti- cle errors in essays written by non-native En- glish speakers; we show that training on data that contain errors produces higher accuracy when compared to a system that is trained on clean native data. We propose several train- ing paradigms with error generation and show that each such paradigm is superior to training a classifier on native data. We also show that the most successful error generation methods are those that use knowledge about the arti- cle distribution and error patterns observed in non-native text. 1 Introduction This paper considers the problem of training clas- sifiers to detect and correct errors in grammar and word usage in text. Both native and non-native speakers make a variety of errors that are not always easy to detect. Consider, for example, the problem of context-sensitive spelling correction (e.g., (Gold- ing and Roth, 1996; Golding and Roth, 1999; Carl- son et al., 2001)). Unlike spelling errors that result in non-words and are easy to detect, context-sensitive spelling correction task involves correcting spelling errors that result in legitimate words, such as confus- ing peace and piece or your and you’re. The typical training paradigm for these context-sensitive ambi- guities is to use text assumed to be error free, replac- ing each target word occurrence (e.g. peace) with a confusion set consisting of, say {peace, piece}, thus generating both positive and negative examples, re- spectively, from the same context. This paper proposes a novel error generation ap- proach to the problem of training classifiers for the purpose of detecting and correcting grammar and usage errors in text. Unlike previous work (e.g., (Sj¨obergh and Knutsson, 2005; Brockett et al., 2006; Foster and Andersen, 2009)), we selectively intro- duce mistakes in an appropriate proportion. In par- ticular, to create training data that closely resemble text with naturally occurring errors, we use error fre- quency information and error distribution statistics obtained from corrected non-native text. We apply the method to the problem of detecting and correct- ing article mistakes made …

Content truncated to 500 words