We are building a text generation module for a decision support system designed to assist physicians in the management of stroke. This module produces multi-paragraph reports on stroke cases stored in the Stroke Data Base or on cases being processed by the decision support system. Analysis of human-generated case reports using Sager’s Linguistic String Parser (LSP) led to a characterization of the stroke sublanguage in terms of four components: a Text Grammar for stroke case reports, a set of Stroke Information Formats, a Relational Lexicon for the stroke sublanguage, and a Linguistic String Grammar for this sublanguage. At this point, we have produced free text by using reverse transformations from our LSP grammar to combine fragments into sentences. Our future goal lies in discovering how to generate good paragraphs, using these components as tools.