1.3 Types in the Data Model

The abstract data model allows for the following principal levels of representation within an NLG system:

1.

Conceptual, i.e. non-linguistic representations constructed by external agents in accordance with their own purposes

2.

Semantic, i.e. linguistically-oriented representations of meaning

3.

Rhetorical, i.e. structures organised for a rhetorical purpose

4.

Document, i.e. structures pertaining to layout and function of a text as a document

5.

Syntactic, i.e. corresponding to conventional syntactic structure

There are certain natural principles of precedence/ordering between these levels which are assumed:

• Everything is eventually derived from some conceptual representation.
• Rhetorical and document structures need to reference semantic and/or conceptual information.
• The final output of the NLG system will be syntactic structures and certain kinds of document structures.

Within each of the levels thus identified, we have found it useful so far to consider the general distinction between abstract and concrete levels of representation. This may not necessarily turn out to be a distinction that works in the same way for all of them, but it reflects the idea that representations at a given level undergo a progression from being originally proposed by a process that is not necessarily an expert on that level through to being finalised by a process that is. Thus, for instance, a text planner may make recommendations about syntactic structure, but it will take a realisation component (with access to a grammar of the language etc.) to flesh those out into an actual structure, ensuring that the constraints of the grammar are satisfied. For each level of representation, there will be legal operations that can be carried out in order to transform the originally proposed representation (the ``abstract'' representation) into the finalised one (the ``concrete'' one). These operations will not necessarily be simple monotonic ones (e.g. a realisation component might be able to change syntactic proposals that are somehow incompatible with any legal syntactic structure).

The following table displays the 10 levels of representation thus identified, with brief notes on our approach to them in RAGS.

Level Abstract Concrete

Conceptual Out of scope Not yet considered

Semantic Possibly = Concrete Conceptual Considered

Rhetorical Considered Considered

Document Considered Out of scope

Syntax Considered Out of scope

Several levels of representation are considered outside the scope of RAGS. Abstract conceptual representations will not be considered in RAGS because they involve decisions made by processes outside an NLG system. Concrete syntactic representations will not be considered because this level of representation is heavily theory-dependent and has been addressed at length in other work. The same applies to concrete document structure, which could be expressed in a number of existing layout/ formatting languages, for instance.

Concrete conceptual representations have not yet been considered (though we have an initial proposal for how an NLG system can interface in general to a knowledge base). There is a possibility that in fact concrete conceptual and abstract semantic representations should be the same (section 9.1 below). This is a major unresolved issue.