Integrating Speech and Natural-Language Processing

INTEGRATING SPEECH AND NATURAL-LANGUAGE PROCESSING Robert Moore, Fernando Pereira, and Hy Murveit SKI International Menlo Park, California 94025 ABSTRACT SRI has developed a new architecture for integrating speech and natural-language processing that applies linguistic constraints during recognition by incrementally expanding the state-transition network embodied in a unification grammar. We compare this dynamic-gralnlnar-network (DGN) approach to its principal alternative, word-lattice parsing, presenting preliminary experimental results that suggest the DGN approach requires much less computation time than word-lattice parsing, while maintaining a very tractable recognition search space. INTRODUCTION We report on an innovative and highly effective architecture for integrating speech and natural-language (NL) processing. This architecture takes full advantage of the linguistic constraints supplied by an NL processor, yet it maintains a very tractable recognition search space and appears to require dramatically less computation by the NL processor than do other approaches. It is useful to begin by recalling why one would want to integrate speech and NL processing. There are two primary reasons why this is desirable. First, there are many applications of spoken-language processing that require not just recognition but understanding, such as spoken-language systems for database query and other computer interface applications. Second, constraints on natural language can be used to reduce the speech recognition search space and therefore improve recognition accuracy. To take an example from our own system, one of the sentences in the 1987 DARPA resource management speaker-independent test set is the following: What is the ETA at her destination of Fanning? Without any source of grammatical constraints (perplexity -- 1000), one version of the SRI system recognizes this sentence as: Why added ETA at her destination of Fanning. This hypothesis contains three errors out of nine words in the original sentence. With the fairly modest degree of grammatical constraint represented by a perplexity-510 natural-language grammar, however, our system was able to recognize this sentence with no errors. The point is that even though the incorrect hypothesis represented a better phonetic and phonological match to the input signal according to the models used by our system, with an NL grammar the system was able to rule out the incorrect hypothesis and choose the correct, although lower-scoring, hypothesis. While this is, of course, a carefully selected example of our system's performance, it is a good illustration of what we are trying to acheive. 243 OTHER APPROACHES TO SPEECH/NL INTEGRATION Prior to the work reported here, there have been two predominant approaches to integrating speech and NL processing. The first is the serial connection of speech and NL processors, and the second is word-lattice parsing. In the serial approach, the speech recognizer makes its best guess as to what word string was uttered, and the NL processor parses and interprets the word string as if it were text. This approach has the advantages that it is simple to implement, and it is relatively fast because neither process increases the search space of the other. Unfortunately, it takes no advantage of the NL constraints embodied in the NL processor …

Content truncated to 500 words