ABSTRACT
Natural language parsers are usually adapted to textual input data, which strictly follows a set of grammatical rules (Sabah 1988). We are here interested in dealing with casual speech, which requires a high degree of robustness. The computational model used for this purpose is a parallel architecture developed in the framework of human–machine interaction, and which is based on a real-time working principle, coincidence detection. Applications of this processing principle to parsing have already been investigated, using different formalisms and focusing either on robustness (Jacquemin 1992) or on learning mechanisms (Roques 1994). The present paper concentrates on a formalism based on Guided Propagation Networks (GPNs) (Béroule 1990), proposing a parser which deals with different kinds of noise after having learnt syntactic structures from examples of correct sentences.
