AGILE Review meeting - April 2000
The AGILE IMD sentence planner
®Discourse aggregation
®Rhetorical relations like Purpose, Means
®Retrievability from context
®Using the discourse model maintained by the text planner
®Style
®Different text styles: personal style, impersonal style
®Expression of location: implicit or explicit side-effects
®Various levels of complexity in lexical semantics
®One-to-one mapping between concepts and words
®One-to-many mapping between concepts and words
®Translation of content structure to lexical-semantic structure
®
Depending on how text plan elements relate, the sentence planner may decide to relate them using a rhetorical relation like Purpose or Means. For example, if a task (say Task 1) is planned as the only possible way to achieve a (higher) task (Task 2), then the sentence planner can decide to generate a sentence plan in which the content related with Task 1 and the content related with Task 2 is aggregated: “Do <Task 1> in order to <Task 2>”

The sentence planner makes use of the discourse model maintained by the text planner in order to indicate, in a sentence plan, whether a particular piece of content should be considered to be retrievable from the already established context or not. Like we said, the Czech grammar uses such information included in the text plan in order to generate a contextually appropriate word order, ensuring a smooth (i.e. coherent) flow in the output text.

The sentence planner also includes the realization constraints (already imposed by the text planner) to ensure that a lexico-grammar generates the content in the appropriate style (impersonal, personal).

If in the content specification a side-effect indicates the appearance of a interface modality (e.g. a dialogbox), but the text should not include explicit side-effects, then the modality is planned as a location with an immediately subsequent task (if possible).

Finally, the sentence planner is capable of dealing with various levels of complexity in lexical semantics – specifying how a particular concept should be linguistically realized. In the easiest case, one concept maps directly onto a single word, like the concept “file” maps directly onto the Czech word “soubor”. However, there are also more complex cases:
•A single concept should be mapped onto a configuration of several words, like the concept “dialogbox” onto the Czech “dialogovy okno”. It is important to be able to map onto a configuration (rather than a fixed string) for these words may have to be inflected, depending on the context in which they will be used.
•An entire conceptual structure (concept+attributes) needs to be translated into a different (non-homomorphic) lexical structure. An example of this is the case of numerals, where we need to alter the attribute structure (kinds of attributes and their interrelation).
The sentence planner has been implemented such that, for a given language, one need only provide a file in which these kinds of mappings are provided – these mappings are not hardwired into the sentence planner module itself.