Grammatical error prediction

In this thesis, we investigate methods for automatic detection, and to some extent correction , of grammatical errors. The evaluation is based on manual error annotation in the Cambridge Learner Corpus ((((), and automatic or semi-automatic annotation of error corpora is one possible application, but the methods are also applicable in other settings, for instance to give learners feedback on their writing or in a proofreading tool used to prepare texts for publication. Apart from the , we use the British National Corpus (((() to get a better model of correct usage, WordNet for semantic relations, other machine-readable dictionaries for or-thography/morphology, and the Robust Accurate Statistical Parsing ((((() system to parse both the and the and thereby identify syntactic relations within the sentence. An ancillary outcome of this is a syntactically annotated version of the , which we have made publicly available. We present a tool called GenERRate, which can be used to introduce errors into a corpus of correct text, and evaluate to what extent the resulting synthetic error corpus can complement or replace a real error corpus. Different methods for detection and correction are investigated, including: sentence-level binary classification based on machine learning over n-grams of words, n-grams of part-of-speech tags and grammatical relations; automatic identification of features which are highly indicative of individual errors; and development of classifiers aimed more specifically at given error types, for instance concord errors based on syntactic structure and collocation errors based on co-occurrence statistics from the , using clustering to deal with data sparseness. We show that such techniques can detect, and sometimes even correct, at least certain error types as well as or better than human annotators. We finally present an annotation experiment in which a human annotator corrects and supplements the automatic annotation, which confirms the high detection/correction accuracy of our system and furthermore shows that such a hybrid setup gives higher-quality annotation with considerably less time and effort expended compared to fully manual annotation. Preface First of all, there are many people who deserve to be mentioned on this page, but whose names do not appear — from my parents, who hardly ever lost their patience when I bombarded them with questions as a child, to the visiting friend who presented me with the idea of a doctorate over dinner at a time when further studies did not seem an obvious choice —, and I should like to …