The Delphi Natural Language Understanding System

The Delphi Natural Language Understanding System Madeleine Bates, Robert Bobrow, Robert Ingria and David Stallard BBN Systems and Technologies, Inc. 70 Fawcett St. Cambridge, MA 02138 Abstract This paper presents Delphi, the natural language component of the BBN Spoken Language System. Delphi is a domain- independent natural language question an- swering system that is solidly based on linguistic principles, yet which is also ro- bust to ungrammatical input. It includes a domain-independent, broad-coverage gram- mar of English. Analysis components include an agenda-based best-first parser and a fall- back component for partial understanding that works by fragment combination. Delphi has been formally evaluated in the ARPA Spoken Language program's ATIS (Airline Travel In- formation System) domain, and has performed well. Delphi has also been ported to a spo- ken language demonstration system in an Air Force Resource Management domain. We dis- cuss results of the evaluation as well as the porting process. 1 Introduction Delphi is a natural language understanding system based on general linguistic principles which is adaptable to any question-answering domain. It incorporates a number of domain-independent knowledge bases, including a gen- eral, broad-coverage grammar of English with a pow- erful and flexible handling of complementation. Unlike most other linguistically motivated systems, however, Delphi is also highly robust, allowing for partial under- standing when an input is ungrammatical, disfluent, or not properly transcribed by a speech recognizer. Thus, Delphi can be used for a spoken language application as readily as for a written one. Furthermore, Delphi's partial understanding component, called the Semantic Linker, is driven off the same system of semantic rules as Delphi's regular best-first parser. Building a robust application therefore requires no additional effort. There are several components of the system, which is diagrammed in Figure 1. First are the parser and Semantic Linker, which output an intermediate repre- sentation we call a "semantic graph". The semantic Z t I ..EL L.J 1 DATABASE ] 1 RESPONSE Figure 1: System Diagram graph is passed to a quantification stage which pro- duces a fully scoped logical form from it. The logical form is then passed to the discourse stage, which re- solves pronominal references and performs other types of task-dependent constraint resolution to produce the final logical form. The final logical form is then passed to the backend translator, and then to the application system which produces the response. Several knowl- edge bases are employed by these analysis components, including grammar, "realization rules" and the domain model, which represents the set of classes and binary relations of the given application domain. Delphi differs from most other linguistically moti- vated systems in the role that is played by syntax. The primary function of Delphi's parser and syntac- tic knowledge bases is not to produce a parse tree, but rather to constrain the search for an appropriate semantic graph interpretation of the utterance. Semantic graphs are produced not by rule-to-rule compositionality, but by what might be called "relation-to-relation" compo- sitionality - the association of grammatical relations in the syntactic structure …

Content truncated to 500 words