Identification of a Writer’s Native Language by Error Analysis

This dissertation is the result of my own work and includes nothing which is the outcome of work done in collaboration except where specifically indicated in the text. This dissertation does not exceed the regulation length of 15, 000 words, including tables and footnotes. Summary In this project, we investigate the task of native language identification. We study a set of Indo-European languages, and demonstrate how machine learning techniques can be used to identify native language of a text's author. A number of different features are extracted and applied to this task. Their contribution to overall performance is investigated and reported. We explore the hypotheses that the choice of words in a free text is influenced by a writer's native language, and that the errors committed by a writer are based on the differences between the writer's native language system and that of English. We identify the error types typical for speakers of different native languages, and show how using different features based on the discriminative error types can improve classification. Acknowledgments I would like to thank my supervisor, Prof. Ted Briscoe, for his guidance and constant support. I am grateful for his encouragement and valuable suggestions throughout the course of this work. I would also like to thank Helen Yannakoudakis and Øistein Andersen for their much appreciated help and their ability to identify the weak spots in my work and offer suggestions for improvement.