TINA: A Natural Language System for Spoken Language Applications

A new natural language system, TINA, has been developed for applications involving spoken language tasks. TINA integrates key ideas from context free grammars, Augmented Transition Networks (ATN's), and the unification concept. TINA provides a seamless interface between syntactic and semantic analysis, and also produces a highly constraining probabilistic language model to improve recognition performance. An initial set of context-free rewrite rules provided by hand is first converted to a network structure. Probability assignments on all arcs in the network are obtained automatically from a set of example sentences. The parser uses a stack decoding search strategy, with a top-down control flow, and includes a feature-passing mechanism to deal with long-distance movement, agreement, and semantic constraints. TINA provides an automatic sentence generation capability that has been effective for identifying overgeneralization problems as well as in producing a word-pair language model for a recognizer. The parser is currently integrated with MIT's SUMMIT recognizer for use in two application domains, with the parser screening recognizer outputs either at the sentential level or to filter partial theories during the active search process.