Name of Faculty Adviser Signature of Faculty Advisor Date

Word sense discrimination is the problem of identifying different contexts that refer to the same meaning of an ambiguous word. For example, given multiple contexts that include the word ’sharp’, we would hope to discriminate between those that refer to an intellectual sharpness versus those that refer to a cutting sharpness. Our methodology is based on the strong contextual hypothesis of Miller and Charles (1991), which states that ”two words are semantically related to the extent that their contextual representations are similar.” This thesis presents corpus–based unsupervised solutions that automatically group together contextually similar instances of a word as observed in a raw text. We do not utilize any manually created or maintained knowledge–rich resources such as dictionaries, thesauri or annotated corpora. As a result, our approach is well suited to the fluid and dynamic nature of word meanings. It is also portable to different domains and languages, and scales easily to larger samples of text. The overall objective of this thesis is to study the effect of various feature types, context representations and clustering methods on the accuracy of sense discrimination. We also apply dimensionality reduction techniques to capture conceptual similarities among the contexts and don’t just rely on the surface forms of words in the text. We present a systematic comparison of various discrimination techniques proposed by Pedersen and Bruce (1997) and Schutze (1998). We find that the first order method of Pedersen and Bruce performs well with larger amounts of text, but that the second order method of Schutze is more effective with smaller data sets. We also discovered that a divisive approach is more suitable for clustering smaller set of contexts, while the agglomerative method performs better on larger data. We conducted experiments to study the effect of using various sources of training, and found that local contexts of a word provide better discrimination features than a running text like complete newspaper articles. We compared the performance of our knowledge–lean method against that of a knowledge– intense approach, and found that although the latter was successful in conjunction with smaller datasets, it didn’t show significant improvements with larger data. This suggests that the features learned from a large sample of text certainly have the potential to outperform those learned from a knowledge-rich resource like dictionary.