The MIT ATIS System: February 1992 Progress Report

T]he MIT ATIS System: February 1992 Progress Report 1 Victor Zue~ James Glass, David Goddeau, David Goodine, Lynette Hirschman, Michael Phillips, Joseph Polifroni, and Stephanie Seneff Spoken Language Systems Group Laboratory for Computer Science Massachusetts Institute of Technology Cambridge, Massachusetts 02139 ABSTRACT This paper describes the status of the MIT ATIS system as of February 1992, focusing especially on the changes made to the SUMMIT recognizer. These include context-dependent phonetic modelling, the use of a bigram language model in conjunction with a probabilistic LR parser, and refinements made to the lexicon. Together with the use of a larger training set, these modifications combined to reduce the speech recog- nition word and sentence error rates by a factor of 2.5 and 1.6, respectively, on the October '91 test set. The weighted error for the entire spoken language system on the same test set is 49.3%. Similar results were also obtained on the February '92 benchmark evaluation. INTRODUCTION This paper presents an update on the MIT ATIS sys- tem, which has been under development since 1990. We will describe several changes made to our system since the last official common evaluation in February, 1991 [8], with particular emphasis on the speech recognition com- ponent.. We will also present our evaluation results for the October '91 "dry-run" test set and the February '92 test set. We have also modified our natural language component to include a robust parsing strategy. This change is described in detail in a companion paper [9]. SPEECH RECOGNITION In this section we will describe the changes we have made over the past year to the speech recognition com- ponent (SUMMIT) of our ATIS system. These include im- provements to both the phonetic and language models, and refinements on the lexicon. We have also imple- mented the acoustic models on a set of DSP boards to allow near real-time evaluation and demonstration. The baseline SUMMIT system uses a mixture of up to 16 diagonal Gaussian models for each lexical unit. In recent months, we have been able to simplify the input representation of the models significantly with no loss in performance. The current representation consists of 39 1This research was supported by i)ARPA under Contract N00014-89-J-1332, monitored through the Office of Naval Research. 84 segmental measurements for each hypothesized segment. This vector is rotated via principal component analysis prior to mixture Gaussian modelling. Segment duration is modelled separately, in the log domain, using a mixture of Gaussians. At the moment, spontaneous disfluencies are represented by one model, and are required to be one segment long. Training and Testing Corpora The multi-site ATIS data collection effort has resulted in a significant increase in the amount of speech data available to the community [6]. For speech recognition system development, we started with all the MADCOW data released by NIST, and augmented them with ATIS data collected earlier at MIT. Some 9,711 utterances in this pool were designated as training material, and an ad- ditional 1,595 utterances were set aside as a development …

Content truncated to 500 words