Portable Natural Language Generation using SPOKESMAN

Portable Natural Language Generation using SPOKESMAN Abstract 1 This paper reports on the SPOKESMAN natural language generation system, which is a domain independent text gen- erator designed to incrementally produce text for an underly- ing application program. This work is a direct outgrowth of the work we reported on at the last ACL Applied Conference in 1988, where we connected an application program directly to the linguistic component, Mumble-86. The major addi- tion has been a new component to the system, a text planner that provides the capability to compose the utterance incre- mentally. The central feature of the text planning com- ponent is a new level of representation that both captures more linguistic generalizations and makes the system more portable, so that we can easily interface to different domains and different kinds of application programs. This larger sys- tem is called "Spokesman", as it acts as the mouthpiece for a number of application programs. 1. Introduction There are generally two approaches to achieving portabil- ity. One is to build peripheral components that can auto- matically or semi-automatically acquire the knowledge needed to port to a new domain. The second is to modular- ize the system so that the components that are not domain specific are kept separate from those that are, and to try to maximize the amount of knowledge in the more general components. In the long term, a combination of these approaches will certainly be needed. In the work presented here, we have concentrated on the later. For example, by taking advantage of the fact that most of language is not par- ticular to any domain, we have isolated the linguistic realiza- tion component and used the Text Structure to capture abstract linguistic generalizations. Also by using the knowledge base of the application directly we can not only capture generalizations about objects which are expressed similarly, but also handle those cases where the means of expression is specific to a particular domain. 2. Modularization In our earlier work, generation involved three modules: a linguistic realization component (LRC) MUMBLE-86 (Meteer, et.al 1987), an underlying application program, and a special purpose text planner to handle the mapping from one to the other. The text planner used the input specifica- tion language to Mumble-86 as a means of compensating 1 The work reported here was done at BBN Systems and Technologies, Cambridge, Massachusetts. I would like to thank all of my BBN colleagues, in particular Glenn Abrett, Stephen Deutsch, Dawn MacLaughlin, and Varda Shaked, for their contributions to this work during my tenure at BBN. Marie Meteer Sage Lab Rensselaer Polytechnic Institute Troy, New York 12180 METEER@CS.RPI.EDU for the semantic deficits of linguistically naive underlying application programs without compromising principled grammatical treatments in natural language generation. V///~/ALBM.spedfic text planner/,//A ('u nderl ying represent ation • I(EE ) Y~/~ iu . . . . . . . . . . . : . . _ . . n :::::::::::::::::::::::::::::::::::::::::::::::::: LANOUAOE ::::::::::::::::::::::::::::::::::::::::::::::::::::::: 1988 Architecture: Interfacing to Mumble-86 While this modularization isolated the …

Content truncated to 500 words