Learning Fault-Tolerant Speech Parsing with SCREEN

This paper describes a new approach and a system SCREEN’ for fault-tolerant speech parsing. Speech parsing describes the syntactic and semantic analysis of spontaneous spoken language. The general approach is based on incremental immediate flat analysis, learning of syntactic and semantic speech parsing, parallel integration of current hypotheses, and the consideration of various forms of speech related errors. The goal for this approach is to explore the parallel interactions between various knowledge sources for learning incremental fault-tolerant speech parsing. This approach is examined in a system SCREEN using various hybrid connectionist techniques. Hybrid connection& techniques are examined because of their promising properties of inherent fault tolerance, learning, gradedness and parallel constraint integration. The input for SCREEN is hypotheses about recognized words of a spoken utterance potentially analyzed by a speech system, the output is hypotheses about the flat syntactic and semantic analysis of the utterance. In this paper we focus on the general approach, the overall architecture, and examples for learning flat syntactic speech parsing. Different from most other speech language architectures SCREEN emphasizes an interactive rather than an autonomous position, learning rather than encoding, flat analysis rather than in-depth analysis, and fault-tolerant processing of phonetic, syntactic and semantic knowledge. Introduction and Motivation In the past, the analysis of spontaneous speech utterances as ‘syntactic and semantic case frame representations received relatively little attention. Although there had been some early attempts for combination (Erman eZ al. 1980) the restricted speech and language techniques at that time forced each field, speech and language processing, to concentrate on developing further techniques separately. Therefore, in the last decade there have been primarily isolated modular attempts to build speech analyzers (e.g., (Lee, Hon, & Reddy 1990; McClelland & Elman ‘SCREEN stands for Symbolic Connectionist Robust EnterprisE for Natural language 670 Machine Learning 1986)) or language analyzers (e.g., (Hobbs et al. 1992; Kitano & Higuchi 1991)). However, recent approaches attempt to integrate speech and language earlier to reduce the extensive space of acoustic, syntactic and semantic hypotheses (Pyka 1992; Young et al. 1989). The MINDS system (Young et al. 1989) is a speech language system which combines a speech recognizer (Lee, Hon, & Reddy 1990) with expectation-driven language analysis. The main contribution of the MINDS system is its early integration of speech hypotheses with language hypotheses in order to restrict the search space for speech processing. On the other hand, the MINDS system relies heavily on hand-coded pragmatic knowledge from a single domain. The ASL system (e.g. (Pyka 1992)) is a speech language system which focused on the examination of interactions in a very general architecture. This system has an architecture similar to a blackboard architecture but without explicit control. Autonomous components can send and receive hypotheses, but the overall architecture and relationships between the components are flexible. While the MINDS system emphasized the use of pragmatic knowledge for supporting speech processing, the ASL system focused rather on syntactic and semantic knowledge. The ASL system has an extremely flexible architecture which can avoid early mistakes in favoring a particular architecture. On the other hand, this flexibility also requires very sophisticated communication operations for complexer interactions. Both MINDS and ASL belong to the state-of-the-art architectures in speech language systems. However, in both systems the language knowledge is basically munuully encoded and domain-dependent. Furthermore, currently errors like false starts, hesitations, corrections, and repetitions have only been implemented in a rudimentary pragmatic manner in the MINDS system. We designed SCREEN as a system for learning fault-tolerant incremental speech parsing. SCREEN deals with repairs (Levelt 1983), false starts, hesitations, and interjections. Since connectionist techniques have inherent fault tolerance and learning capabilities we explore these properties in a hybrid connectionist From: AAAI-94 Proceedings. Copyright © 1994, AAAI (www.aaai.org). All rights reserved.