This paper addresses the area of text generation known as microplanning [Levelt1989,Panaget1994,Huang and Fiedler1996], or sentence planning [Rambow and Korelsky1992]; [Wanner and Hovy1996]. Microplanning involves low-level discourse structuring and marking, sentence boundary planning, clause internal structuring and all of the varied subtasks involved in lexical choice. Conventional wisdom dictates that these complex tasks be modularized and treated separately:
Since sentence planning tasks are not single-step operations, since they do not have to be performed in strict sequence, and since the planner's operation is non-deterministic, this suggests that each sentence planning task should be implemented by a separate module or by several modules. [Wanner and Hovy1996]
Such an argument is natural if generation is viewed as a set of coarse-grained tasks. Indeed, with the exception of a few researchers ([Elhadad et al. 1997] and the incrementalists listed below), the task-oriented view is standard in the generation community. Unfortunately, task-oriented generation sets up unnecessary barriers among the components of the generation process, primarily because, in a realistic scenario, the tasks are intertwined to a high degree. Overcoming these barriers has become a central topic in generation research; however, the very wording of the argument quoted above --- ``not performed in a strict sequence,'' ``non-deterministic'' --- suggests to us that this approach is not a natural one. In our approach the basis of modularization is sought in the nature of the input data to the generation process, in our case, a text meaning representation, formulated largely in terms of an ontology. This data-oriented approach is similar to that taken by many incremental generators [De Smedt1990,Reithinger1992], although these tend to concentrate on syntactic processing. But see [Kilger1997], who explicitly addresses microplanning. We feel that our work provides an optimal path between task-oriented generators (which face problems due to the interrelationships between the tasks) and traditional incremental generation (which does not take advantage of problem decomposition as discussed below).
In what follows we describe our ontology-based modularization, the kind of constraints which can be automatically set up within such a paradigm, and the control mechanism we employ to process it. We conclude with a discussion of the avoidable barriers inherent in most current approaches, along with their attempts at circumventing them, and how our approach eliminates many of the problems. We also point out differences between our approach and that of the incremental generators.