The fundamental role of pragmatics in Natural Language Understanding and its implications for modular, cognitively motivated architectures

After having compared four methods of studying language, I will examine their respective consequences with regard to architecture natural language understanding and generation. I will concentrate in this paper mainly on two points: how and when to use pragmatic knowledge for automatic text understanding. I will then show how to overcome the inadequacies of the control mechanisms traditionally used in Natural Language Understanding systems, and the usefulness of multi-expert systems for maintaining modularity without introducing artificial ambiguities. In addition, I propose some extensions to the classical reflectivity in order to make natural language understanding systems reflective. Taking all these factors into account led to CARAMEL-1: a general natural language understanding system, which was meant to be general enough in order to apply to various kinds of applications (The acronym CARAMEL stands for “Compréhension Automatique de Récits, Apprentissage et Modélisation des Échanges Langagiers” meaning roughly “Automatic Story Understanding, Learning and Dialogue Management”). One of the problems that arise in a large modular architecture is to control explicitly all the processes. However, this is not only cumbersome, but also it does not account for automatic reflex processes, yet these latter are essential for natural language understanding. This is why I have introduced a new kind of memory model (the Sketchboard) which implies not only various kinds of relations between the processes but allows also for reactive feedback loops. 1. LIMSI B.P. 133, 91403 ORSAY CEDEX — FRANCE Tel: 33 (1) 69 85 80 03, Fax: 33 (1) 69 85 80 88, E-mail: [email protected]