A Tensor Encoding Model of Word Meaning Theory and Application to Information Retrieval A THESIS SUBMITTED TO THE SCIENCE AND ENGINEERING FACULTY OF QUEENSLAND UNIVERSITY OF TECHNOLOGY IN FULFILMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY

The ability to create meaning from our environment is fundamental to the way in which we interact with our world. Research into human cognition and language has inspired the development of a class of computational models that induce meaning based on how words are associated in streams of natural language. These models represent semantic associations between words within a geometric space, hence their popular name, semantic space models. Their low cost and ability to provide effective performance on a number of linguistic tasks, including synonym judgement and analogy making, has seen them promoted as having potentially transformative powers for improving communication between humans and computers. This thesis develops a formal model of word meaning, that extends recent advances in semantic space technology, including the use of tensor representations, and evaluates its performance on a number of semantic tasks, including synonym judgement, semantic distance, semantic categorization and similarity judgement of medical concepts. The model’s formal framework, combined with its superior efficiency and effectiveness compared to existing models, provides a theoretical platform from which it is applied to the query expansion process within the information retrieval problem. Its effectiveness in this role is evaluated on several ad hoc retrieval and web search tasks. The demonstrated significant improvements in retrieval effectiveness of this approach over benchmark models motivates a discussion of other potential applications and areas for future work.