Domain Communication Knowledge

This paper advances the hypothesis that any text planning task relies, explicitly or implicitly, on domainspecific text planning knowledge. This knowledge, "domain communication knowledge", is different from both domain knowledge and general knowledge about communication. The paper presents the text generation system Joyce, which represents such knowledge explicitly. The Joyce Text Generation System The Joyce text generation system is a part of the software design environment Ulysses (Korelsky and Ulysses Staff, 1988; Rosenthal el al., 1988). Ulysses is a graphical environment for the design of secure, distributed software systems. The design proceeds hierarchically by top-down refinement. A formal specification interface and a theorem prover allow the user to formally verify the flow security of the designed system. Joyce is part of the user interface. Joyce generates different types of texts about software designs: • It generates annotations of the design which are intended to serve as system documentation during and after the design process. • It is used to explain the result of a heuristic security design tool, the "flow analyzer". The texts Joyce generates are specifically conceived of as written texts: there is no feature for interactive natural-language explanation. The texts may be several paragraphs long. The text in figure 1 is an annotation of the component "Host"; the graphical representation of the first level of the design of the Host is shown in figure 2. (This picture corresponds to the first of the two paragraphs of the text.) The text annotates the software design by describing its structure and interpreting it in terms of its security characteristics. *The research reported in this paper was carried out while the author was at Odyssey Research Associates, Ithaca, NY. It was supported by the Air Force Systems Command at Rome Air Development Center under Contract No. F30602-85-C-0098 S t r u c t u r e o f Joyce Joyce consists of three separate modules, which perform distinct tasks and access their own knowledge bases. 1. The text planner produces a list of propositions, which represents both the content and the structure of the intended text. Thus, the task of Joyce's text planner is similar in definition to TEXT's (McKeown, 1985), but different from that of Penman (Hovy, 1988), which expects the content selection task to have already been performed. Each proposition is expressed in a language-independent, conceptual framelike formalism. It encodes a minimal amount of information, but can be realized as an independent sentence if necessary. The text planner draws on domain communication knowledge expressed in a high-level schema language (see below). 2. The sentence planner takes the list of propositions and determines how to express them in natural language. This task includes choosing lexicalizations and a syntactic structure for each propositions, and assembling these lexico-syntactic structures into larger sentences. It draws on knowledge captured in the conceptual/Engish dictionary. 3. The linguistic realizer takes the syntactic structures and produces surface sentences. It draws on syntactic and morphological knowledge, expressed in the lexicon. The linguistic component is based on MeaningText Theory (Mel'~uk, 1988), and is a reimplementation in Lisp of Polgu~re's Prolog implementation (see (Iordanskaja et al., 1988a; Iordanskaja et at., 1988b)). Usually, the task of text generation is subdivided into two subtasks (planning and realization), not three. However, there is a certain amount of disagreement about where the line between the two is to be drawn. For example, McKeown's TEXT (McKeown, 1985) and R/Ssner's SEMTEX (RSsner, 1987) seem to consider the tasks that Joyce classifies as sentence planning as part of the realization process, whereas Meteer's SPOKESMAN (Meteer, 1989) classifies them as part of text planning. The proposed finer-grained terminology may prove useful in discussing text generation and