Stochastic versus rule-based speech understanding for information retrieval

Abstract

In this paper we report our experience at LIMSI-CNRS in developing and porting a stochastic component for natural language understanding to different tasks and human languages. The domains in which we test this component are the American ATIS (Air Travel Information Services) and the French MASK (Multimodal-Multimedia Automated Service Kiosk) applications. The study demonstrates that for limited applications, a stochastic method outperforms a well-tuned rule-based component. In addition we show that the human effort can be limited to the task of data labeling, which is much simpler than the design, maintenance and extension of the grammar rules. Since a stochastic method automatically learns the semantic formalism through an analysis of these data, it is comparatively flexible and robust.