In [Vaughan and McDonald1986] and [Meteer1990] a model of text revision in natural language generation is proposed that is based on an integrated approach to parsing and generation. It is based on the observation that during composition of a text, a multi-pass system of writing and rewriting is used to produce an optimal text. They describe a model that includes a generator, a parser and an evaluation component which assesses the parse of what the generator had produced and determines strategies for improvement. The revision process they outline is modelled in terms of the following three phases:
The recognition phase is responsible for parsing text and building a representation rich enough to be evaluated in terms of how well the text coheres. The text representation they developed is called text structure and serves as intermediate level of representation of text planning (cf. [Meteer1990]). The recognition phase analyzes the text as it proceeds using a set of evaluation criteria (e.g., to find ambiguous referents, flag places where optimizations may be possible, such as predicate nominal). For each problem there is a set of one or more strategies for correcting it. The task of the editing phase is to determine which of these strategies to employ (e.g., if the subject has a relation to a previous referent which is not explicitly mentioned in the text, more information may be added through modification). The final step is actually making the change once the strategy has been chosen. This essentially involves marking the input to the generator, so that it will query the revision component at appropriate decision points. For example, if the goal is to put two sentences into parallel structure, the generator when reaching this marker asks the revision component whether it should realize two main clauses or if it should realize one as a subordinate and how it should be realized (e.g., active or passive).