Cluster-Based Pattern Recognition in Natural Language Text

ii Acknowledgements I would like to thank my adviser Prof. Tishby for his guidance and assistance in producing this work, and for his suggestions and positive input. I would also like to thank Beata Beigman Klebanov for her constant help and advice throughout this work, including (but definitely not limited to) the contribution of the parsed data used here. Also deserving of thanks are my family, for their support, and especially my grandmother, Rose Brody, for her confidence in my achievements. iii Abstract This work presents the Clustered Clause structure, which uses information-based clustering and dependencies between sentence components to provide a simplified and generalized model of a grammatical clause. We show that this representation, which is based on dependencies within the sentence, enables us to detect complex textual relations at a higher level of context. The relations we detect are of interest in themselves, as linguistic phenomena, and are also highly suited for use in certain linguistic and cognitive tasks. We define and search for several types of patterns, moving from basic patterns to more complex ones, from patterns within the sentence to those involving entire sentences. Examples of recognized patterns of each type are presented, and also descriptions of several interesting phenomena detected by our method. We assess the quality of the results, and demonstrate the importance of the clustering and dependency model we chose. The principles behind our method are largely domain-independent, and can therefore be applied to other forms of structured sequential data as well.