Using Focus to Generate Complex and Simple Sentences

Using Focus to Generate Complex and Simple Sentences Marcia A. Derr Kathleen R. McKeown AT&T Bell Laboratories Murray Hill, NJ 07974 USA and Department of Computer Science Columbia University Department of Computer Science Columbia University New York, NY 10027 USA Abstract One problem for the generation of natural language text is determining when to use a sequence of simple sentences and when a single complex one is more appropriate. In this paper, we show how focus of attention is one factor that influences this decision and describe its implementation in a system that generates explanations for a student advisor expert system. The implementation uses tests on functional information such as focus of attention within the Prolog definite clause grammar formalism to determine when to use complex sentences, resulting in an efficient generator that has the same benefits as a functional grammar system. 1. Introduction Two problems in natural language generation are deciding what to say and how to say it. This paper addresses issues in the second of these tasks, that of surface generation. Given a semantic representation of what to say, a surface generator must construct an appropriate surface structure taking into consideration a wide variety of alternatives. When a generator is used to produce text and not just single sentences, one decision it must make is whether to use a sequence of simple sentences or a single complex one. We show how focus of attention can be used as the basis on which this decision can be made. A second goal of this paper is to introduce a formalism for surface generation that uses aspects of Kay's functional grammar (Kay, 1979) within a Prolog definite clause grammar (Pereira and Warren, 1980). This formalism was used to implement a surface generator that makes choices about sentence complexity based on shifts in focus of attention. The implementation was done as part of an explanation facility for a student advisor expert system being developed at Columbia University. 2. Language Generation Model In our model of natural language generation, we assume that the task of generating a response can be divided into two stages: determining the semantic content of the response and choosing a surface structure I. One component makes decisions about which information to include in the response and passes this information to a surface generator. For example, an expert system explanation facility may select part of the goal tree, a particular goal and its antecedent subgoals, to explain a behavior of the system. In the advisor system, the output of this component consists of one or more logical propositions where each proposition consists of a predicate relating a group of arguments. The output includes functional information, such as focus, and some syntactic features, such as number and tense, for convenience. Other information, such as the relationships between propositions, is implicit in the organizational structure of the output. The output of the semantic component is passed on to another component, the surface generator. The job of generator is to use whatever …

Content truncated to 500 words