Learning to solve PP-attachment ambiguities in natural language processing through neural networks

Abstract

A technique is proposed, based on neural networks for dealing with a particular problem of syntactical ambiguity in the process of building the syntactical tree of an Italian sentence, namely the PP-attachment problem. The neural network was used as a daemon for a top-down parser, when it faced multiple entries in the parsing table. The network was trained to solve PP-attachment ambiguities by the well known algorithm for error back-propagation. What is new is the knowledge representation technique in the network, which has been designed to represent the relevant pieces of information about the constituents of the sentence. Performance results are reported and discussed, together with future perspectives.<>