AN EXAMPLE-BASED APPROACH TO PREPOSITIONAL PHRASE ATTACHMENT IN NLP

Abstract

This thesis presents a joint project between the Department of Information Systems & Computer Science (ISCS) and Institute of Systems Science (ISS), both at National University of Singapore. Implemented as a dependent module of Tapestry which is a set of NLP tools developed at ISS, this project aims at disambiguating the re-attachment problems in Natural Language Processing (NLP) in order to improve the performance of syntactic structural analysis of Tapestry. In this project, an Example-Based approach was proposed to solve the reattachment problems, particularly the prepositional phrase attachment problem in NLP. Examples were derived in MIR (Message Intermediate Representation) and presented at the semantic level. As a directed conceptual graph structure generated by Tapestry, MIR gives the semantic concepts of words in a sentence and their relationships. Some relations and the lexical units in MIR were applied as examples to disambiguate the re-attachment problems. The synonym sets of the lexical units in the input sentences arc employed to match the examples in case the lexical units are not in the example database. Worc!Net is used as an on-line thesaurus to obtain the synonym sets of input lexical units. A scheme was proposed to calculate the semantic distance between a lexical unit and its synonym sets. By comparing the semantic distance with the objects to which the input lexical unit may be attached, the attachment is disambiguated. So far I have applied this Example-based approach to the prepositional phrase II attachment and verbal phrase attachment. Promising results were obtained in our experiments. To assess the performance of this approach, a comparison with other approaches of attachment disambiguation was made in this thesis. Some advantages of this approach were also analyzed. Moreover, several methods of improving the performance of this Example-based approach were discussed in the last chapter. I envisage that this Example-based framework is also applicable to other attachment problems so that a consistent and general way could be found to handle more kinds of the attachment problems in NLP.