ABSTRACT
In this paper, an alternative architecture is described for a subsymbolic hybrid NLP system that generates abstracts of texts. Traditional AI symbolic processors, namely morphological, syntactic, semantic and pragmatic analysis and synthesis modules (Section 4.4) are co-ordinated with a standard back-propagation Artificial Neural Network (ANN). The novelty of this network lies in its combination of selected information from all these standard modules and using it for pragmatic decisions. More specifically, it assigns degrees of importance to clauses in a text, as a first step towards automated abstract generation. Those with the higher scores are the ones that will be used in the summarization of the text. The high-scoring clauses are subsequently processed by symbolic generation modules which output the abstract, after having performed a number of lexical, syntactic and pragmatic operations on these clauses. The ANN and the related training and testing experiments are described later on in the paper (Section 4.5).
