In [KKR91], Kittredge et al. discuss the notion of domain communication knowledge (DCK for short). They argue that DCK is a third level of knowledge that a NL generator should encode, intermediate between domain knowledge and communication knowledge: DCK is not needed to reason about the domain, but rather, to communicate about the domain. To characterize the difference between DCK on the one hand, and communication knowledge and domain knowledge on the other, they provide the following arguments.
Communication knowledge is the knowledge embodied in the schemata or plan operators that we discussed earlier. [KKR91] illustrate the difference between DCK and communication knowledge by the problem of planning object descriptions:
Consider the task of describing a set of objects in some domain. Communication knowledge about thematic structure implies a strategy that describes together those objects that share some feature. Domain knowledge can supply information about which objects share which features. But if there are many different features, the task remains of choosing the feature(s) according to which the descriptions will be grouped together. This choice must be based on knowledge that is neither general knowledge about communication (since the choice depends on the particular features of objects in the domain) nor actual domain knowledge (since it is only needed for planning communication).
As the difference between DCK and domain knowledge is concerned, [KKR91] discuss the difference between a police crime report and a detective novel.
The underlying domain knowledge is knowledge about events related to the crime, along with general knowledge about human beliefs, desires, intentions and actions. Nonetheless, the two texts generated from this knowledge are quite different. Clearly, the domain knowledge itself does not imply a particular way of communicating about itself.
In the rest of their paper, Kittredge et al. discuss how DCK comes into play in specific kinds of reports, one regarding marine weather forecasts, the other regarding employment statistics. For example, in both cases the order in which the information is presented is fixed. In the weather reports, warnings about bad weather precede the main body of the report. The reports about employment statistics are composed by an overview, followed by multi-paragraph blocks about employment and unemployment respectively. In turn, these two blocks are subdivided into a summary paragraph, followed by paragraphs devoted to changes in (un)employment organized around various parameters (e.g. by sex and by age).
One drawback of the [KKR91] paper is that there is no discussion of how DCK could be expressed in practice. Some of Sibun's local strategies [Sib92] are an example of an explicit encoding of DCK.