PP-attachment disambiguation in natural language processing through neural networks

A special version of the multilayer perceptron is proposed as a suitable oracle for solving meaningful instances of the PP-attacment problem in Italian sentences. The leading idea is to instill into the network those pieces of the syntactical knowledge about the sentence, which can be univocally drawn by a LL(1) parsing algorithm, committing the network to learn to answer «yes» or «no» to the uncertain attachment proposals. This has demanded for a special dynamical architecture of the net, strongly dependent on the run time parsing tree and on the semantics of the words, and for a special representation of the input in order to find out the syntactical invariants throurugh sentences. The learning algorithm is the usual backpropagation algorithm, which turned out to score high percentages of successful attachment. The system is a first building block of an automatic system for the recognition of correctly written sentences.