Grammatical error correction in non-native English

Grammatical error correction (GEC) is the task of automatically correcting grammatical errors in written text. Previous research has mainly focussed on individual error types and current commercial proofreading tools only target limited error types. As sentences produced by learners may contain multiple errors of different types, a practical error correction system should be able to detect and correct all errors. In this thesis, we investigate GEC for learners of English as a Second Language (ESL). Specifically, we treat GEC as a translation task from incorrect into correct English, explore new models for developing end-to-end GEC systems for all error types, study system performance for each error type, and examine model generali-sation to different corpora. First, we apply Statistical Machine Translation (SMT) to GEC and prove that it can form the basis of a competitive all-errors GEC system. We implement an SMT-based GEC system which contributes to our winning system submitted to a shared task in 2014. Next, we propose a ranking model to re-rank correction candidates generated by an SMT-based GEC system. This model introduces new linguistic information and we show that it improves correction quality. Finally, we present the first study using Neural Machine Translation (NMT) for GEC. We demonstrate that NMT can be successfully applied to GEC and help capture new errors missed by an SMT-based GEC system. While we focus on GEC for English, our methods presented in this thesis can be easily applied to any language. Acknowledgements First and foremost, I owe a huge debt of gratitude to my supervisor, Ted Briscoe, who has patiently guided me through my PhD and always been very helpful, understanding and supportive. I cannot thank him enough for providing me with opportunities that helped me grow as a researcher and a critical thinker. I am immensely grateful to my examiners, Paula Buttery and Stephen Pulman, for their thorough reading of my thesis, their valuable comments and an enjoyable viva. My appreciation extends to my fellow members of the Natural Language and Information Processing research group, with whom I have always enjoyed discussing our work and other random things. My gratitude goes to Stephen Clark and Ann Copestake for giving me early feedback on my work as well as Christopher Bryant for generously reading my thesis draft. I would especially like to thank Mariano Felice for being not just a great colleague but also a dear friend. A special mention has …