In existing natural language generation systems, the process
of text generation is divided into three main stages: text
planning, sentence planning, and surface realization
[Rei94], as in Figure 
.
Figure: The stages of language generation
This section briefly describes these tasks, and the issues that arise in performing them in describing the execution of processes. The key point is that to produce the kinds of clear and understandable texts that people need requires rich models of both the causal relationships in the domain, and the inferential capacities of the reader or hearer to reason about those relationships. Although this information must be given explicitly to a natural language generation system, it generally consists of ``obvious'' commonsense facts, and is accordingly painstaking and uninteresting for users to construct. We therefore outline some research strategies for allowing Pat-net designers to specify needed knowledge as naturally and as unobtrusively as possible during programming.