Progress Report on the Chronus System: ATIS Benchmark Results

Progress Report on the Chronus System: ATIS Benchmark Results Roberto Pieraccini, Evelyne Tzoukermann, Zakhar Gorelov, Esther Levin, Chin-Hui Lee, Jean-Luc Gauvain AT&T Bell Laboratories 600 Mountain Avenue Murray Hill, NJ 07974 ABSTRACT The speech understanding system we propose in this paper is based on the stochastic modeling of a sentence as a sequence of elemental units that represent its meaning. According to this paradigm, the original meaning of a sentence, can be decoded using a dynamic programming algorithm, although the small amount of training data currently available sug- gested the integration of the decoder with a more traditional technique. However, the advantage of this method consists in the development of a framework in which a closed training loop reduces the amount of human supervision in the design phase of the understanding component. The results reported here for the February 1992 DARPA ATIS test are extremely promising, considering the small amount of hand tuning the system required. 1. INTRODUCTION In February 1991 [1] we proposed a novel paradigm that represents the conceptual content of a spoken sentence in terms of a probabilistic finite state automaton. The motivations for developing this model are summarized by the following points. • Current natural language understanding systems are generally based on the synthesis of rules oper- ated generally by an expert. This procedure makes maintenance, updating, and generalization of a sys- tem to other tasks a very expensive and difficult op- eration. We bel:eve that an understanding system should incorporate a mechanism that allows, or is suitable for, unsupervised learning. Only using this mechanism can the system easily take advantage of large amounts of training data. • Systems that are based on parsing with formal grammars (finite state, context free, etc) are gener- ally very rigid. A system that has to be integrated with a speech recognizer has to be quite insensitive to recognition errors (substitution, insertion, dele- tion of words) as well as to speech disfluencies, like false starts, ungrammatical sentences, non-speech phenomena, and so on. • The understanding model should define a frame- work that allows an easy and natural integration with the speech recognizer. Following these considerations we formalized the speech understanding problem in terms of a communication channel whose input is the meaning of a sentence and whose output is a sequence of acoustic observations. Here we assume that the meaning of a sentence can be expressed by a sequence of basic meaning units M = ml,m2,...m~VMM and that there is a sequential correspondence between each mj and a subsequence of the acoustic observation A = al,a2...a~v,. This hy- pothesis, although very restrictive, was successfully in- troduced also in [2]. According to this model of the spoken sentence production, one can think of decod- ing the original sequence of meaning units directly from the acoustic observation. The decoding process can be based on the maximization of the a posteriori probabil- ity P(M [ A). The problem now consists in defining a suitable repre- sentation of the meaning of a sentence in …

Content truncated to 500 words