Abstract
This work is a study of the problem of disambiguation of the often confused English locational prepositions, in, at, and on, and the leveraging of semantic knowledge in an automatic way to aid in solving this problem. Although native speakers of a language normally may not find any difficulty with the usage of prepositions, second language learners commonly find them troublesome and often err in their usage of them. This particular set of prepositions is interesting as the three prepositions are commonly mixed-up among second language speakers of English with Chinese as their first language and the prepositions can often be translated as or have a similar function to a single word, the Chinese word在 (zài). We use a decision list, a technique well-known to have been successfully applied to word sense disambiguation to this problem. Then, using a simple, yet effective method that does not require word sense disambiguation, we use WordNet to make overlapping, non-independent word class features available to the decision list learner. This helps us to achieve two goals: 1) we find that adding word class features significantly improve performance to as good as or better than doubling the amount of training data and 2) the combination of the decision list technique and word classes avoids the opaqueness of other black box-type machine learning methods, learning human-interpretable rules that are similar in quality to those described by grammarians.