Third International Conference on Spoken Language Processing (ICSLP 94)
To integrate well into a speech understanding system, a parser should be evidential. The current best evidential approach, stochastic context-free grammar parsing, is not completely satisfactory. This paper presents a new model. Here the parser's input is the raw lattice of word hypotheses, and its output is a set of scored clues to a semantic interpretation; and conversely for feedback. The parser is based on "construction hypotheses" each of which maps directly to word hypotheses, on the one hand, and conceptual hypotheses, on the other. These construction hypotheses are mutually independent, which simplifies the parser. These ideas are incorporated in a tiny prototype speech understanding system.
Bibliographic reference. Ward, Nigel (1994): "A lightweight parser for speech understanding", In ICSLP-1994, 783-786.