Prepositional Phrase Attachment through Semantic Association using Connectionist Approach
Srinivas Medimi, Pushpak Bhattacharyya · 2006
Determining the correct attachment site for Prepositional Phrase (PP) is one of the major sources of ambiguity in natural language parsing and analysis. In this paper, we describe a neural network based approach to prepositional phrase attachment for natural language text. Our approach disambiguates the attachment site for PP through semantic association among the constituents namely verb, noun and PP, using the WordNet semantic classes. It is essentially a corpus based approach. In most of previous corpus based statistical approaches, accurate estimation of probabilities was dependent on the data sufficiency in terms of size and coverage of the features. Moreover, rule-based systems are inappropriate for handling uncertain knowledge. Managing and maintaining rule based systems is also very difficult task and poses many problems. Our method, using the semantic class properties of words, reduces the lexical (word) level data sparseness problem. Neural networks are also very good in capturing the complex nature of semantic association among the words, and as a result capture the selectional restrictions. We have tested our method on Wall Street Journal corpus, and the experimental results show much better accuracy in PP attachment disambiguation and comparable to state-of-the-art approaches and the accuracy of the results shows the effectiveness of our approach.