Computational and Psycholinguistic Approaches to Structural Ambiguity: The Case of Garden Path Sentences

By approaching the phenomenon of structural ambiguity from two different perspectives – psycholinguistic and computational, this thesis shows how linguistic research can have practical applications in improving NLP systems. The analyses of specific sentences show breakdowns in processing, present the possible explanations for the reasons behind them through principles such as Minimal Attachment/Late closure and Lexical preference, and demonstrate the backtracking needed in order to achieve a full, successful parsing of ambiguous sentences. The types of sentences chosen for this thesis are called garden-path sentences, which induce a lot of difficulty in processing for both humans and machines, making them a perfect choice to demonstrate the similarities and differences between sentence processing in humans and machines. The research employs a combination of computational methods, both rule-based and statistical, and psycholinguistic hypotheses to explain the parsing process of garden-path sentences. The results show that in order for an NLP system to fully process highly ambiguous sentences such as these, it needs theoretical input to repair the partly parsed structures and successfully complete the parsing process. This implies a strong need for multidisciplinary research involving programmers, linguists, and cognitive scientists to succeed in emulating human intelligence in complex AI systems