Interacting in natural language with artificial systems: the DONAU project

Abstract This paper is intended to propose a new methodological approach to the conception and development of natural language understanding systems. This new contribution is supported by the design, implementation, and experimentation of DONAU: a general purpose domain oriented natural language understanding system developed and presently running at the Milan Polytechnic Artificial Intelligence Project. The system is based on a two level modular architecture intended to overcome the lack of flexibility and generality often pointed out in many existing systems, and to facilitate the exchange of results and actual experiences between different projects. The horizontal level allows an independent and parallel development of the single segments of the system (syntactic analyser, semantic analyser, information extractor, legality controller). The vertical level ensures the possibility of changing (enlarging or redefining) the definition of the semantic domain on which each particular version of the system is oriented and specialized in a simple, incremental, and user-oriented way. In the paper the general architecture of the system and the mode of operation of each segment are illustrated in detail. Linguistic models, knowledge representation, and parsing algorithms are described and illustrated by means of selected examples. Performance evaluations of the system in the application version on data base inquiry are reported and discussed. Promising directions for future research are presented in the conclusions.