An expert critiquing system differs from most first-generation systems in that it allows the user to suggest his own solution to a problem and then receive expert feedback (the critique) on his proposals. A critique may be presented in different ways - textually, graphically, in tabular form, or a combination of these. In this report we discuss textual presentation. A generalized approach to text generation is presented, in particular in producing the deep structure of a critiquing text.

The generation of natural language falls into two generally accepted phases: deep generation and surface generation. Deep generation involves selecting the content of the text and the level of detail to be included in the text, ie. deciding what to say and how much information to include. Surface generation involves choosing the words and phrases to express the content determined by the deep generator. In this report we discuss the deep generation of a critique.

We present expert critiquing systems and the results of an implementation. Then we review recent advances in text generation which suggest more generalized approaches to the production of texts and we examine how they can be applied to the construction of a critique.

Central considerations in the deep generation of a text involve establishing the goals the text is to achieve (eg. provide the user with the necessary information on which to base a decision), determining a level of detail of such information to be included in the text and organizing the various parts of the text to form a cohesive unit. We discuss the use of Speech Act Theory as means of expressing the goals of the text, the user model to influence the level of detail and the use of Rhetorical Structure Theory for the organization of the text. Initial results from the text organization module are presented.