PP Attachment Ambiguity Resolution with Corpus-Based Pattern Distributions and Lexical Signaturese

Abstract

We propose a method mixing unsupervised learning of lexical pattern frequencies with semantic information which aims at improving the resolution of PP attachment ambiguity. Using the output of a robust parser, i.e. the set of all possible attachments for a given sentence, we query the Web and obtain statistical information about the frequencies of the attachments distributions as well as lexical signatures of the terms on the patterns. All this information is used to weight the dependencies yielded by the parser and eventually to choose of the most probable attachment.