| The deep semantic understanding necessary for complex natural language processing tasks, such as automatic question-answering or text summarization, would benefit from highly accurate word sense disambiguation (WSD). This dissertation investigates what makes an appropriate and effective sense inventory for WSD. Drawing on theories and methodologies from cognitive linguistics, psycholinguistics, and computational linguistics, we address underlying assumptions about sense inventories, investigate factors influencing successful sense annotation, and experiment with the appropriateness of using a lexicon enriched with syntactic and semantic information as a sense inventory for WSD.;A psycholinguistic experiment is presented that investigates how people process polysemous words. The significant facilitation found in processing related senses as compared to unrelated senses supports a model of semantic representation in which related senses are not discrete entities. The results suggest that fine-grained sense inventories may not be necessary for WSD, as little information may be lost from using more general senses. It also suggests that flexible sense inventories, in which systems can move between more or less nuanced senses, may improve the portability of WSD systems.;A series of annotation experiments investigates the factors that affect the reliability of word sense annotation, an important question given the strong correlation between WSD system accuracy and the reliability of the annotation used for training. We found that sense granularity is a significant factor, but that the number of senses used to label a word is not.;Finally, we demonstrate that supervised machine learning can be successful with a sense inventory enriched with syntactic and semantic information, such as VerbNet, thereby consolidating the steps needed to create a complex knowledge representation and reasoning system. We trained a classifier to disambiguate a set of verbs with multiple VerbNet class memberships with 90% accuracy, which represents a 61% error reduction over the baseline. |
