The BBN Spoken Language System

THE BBN SPOKEN LANGUAGE SYSTEM Sean Boisen Yen-Lu Chow Andrew Haas Robert Ingria Salim Roukos BBN Systems and Technologies Corporation 10 Moulton Street Mailstop 009 Cambridge, MA 02238 David Stallard ABSTRACT We describe HARC, a system for speech understand- ing that integrates speech recognition techniques with natural language processing. The integrated system uses statistical pattern recognition to build a lattice of potential words in the input speech. This word lattice is passed to a unification parser to derive all possible associated syntactic structures for these words. The resulting parse structures are passed to a multi-level semantics component for interpretation. INTRODUCTION HARC, the BBN Spoken Language System (Boisen, et al. (1989)) is a system for speech understanding that integrates speech recognition techniques with natural language processing. As our integration methodology, we use lallice parsing. In this architec- ture, an acoustic processor produces a lattice of pos- sible words that is passed to a parser which produces all possible parses for all syntactically permissible word sequences present in the lattice. These parse trees are then passed to a semantic interpretation component, which produces the possible interpreta- tions of these parse structures, filtering out anomalous readings where possible. THE ARCHITECTURE OF HARC In this section, we present a more detailed outline of the general architecture of HARC: 1. An acoustic processor, which uses context- dependent Hidden Markov Models (HMMs) for acoustic modelling, produces a lattice of possible words in the input speech. 2. A chart parser uses a unification grammar to parse the word lattice and produces the set of all possible parses for all syntactically per- missible word sequences in the lattice. The resulting parses are ranked by acoustic likelihood score. 3. A multi-level semantics component processes the parse trees? This component 1This architecture and the names of the associated language levels are from the PHLIQA1 system (Bronnenberg et al. 1980). uses 4 translation steps to derive the mean- ing of each parse. a. The parse tree is converted to an expres- sion of EFL (English-oriented Formal Language); at this level, each word has one EFL constant; this includes words withmultiple senses. b. Each EFL expression is translated into one or more expressions of WML (World Model Language). Where possible, am- biguous constants from an EFL expres- sion are disambiguated and logically equivalent EFL expressions are col- lapsed. c. Each WML expression is converted to an expression in DBL (Data Base Language), which contains one constant for each file in the data base. d. The value of each DBL expression is computed by evaluating the expression against the database; this value is ex- pressed in CVL (Canonical Value Language). For speech understanding, semantics identifies the highest scoring "meaningful" sentence. This sen- tence is the recognized spoken utterance and its meaning is the sytem's interpretation of the input. TRAINING AND TEST SETS To measure the coverage of the syntactic and seman- tic components and the speech understanding perfor- mance of the integrated system, we use the DARPA 1000-word Resource Management Database corpus. This …

Content truncated to 500 words