Julia E. Heine
Repair Strategies for German
Generation in TDMT
- Problems and Prospects -
Abstract:The TDMT generation component has as its input a list of words grouped and
annotated with the help of linguistic markers. These linguistic markers in general
indicate the grammatical function of the respective words, and in theory the task of
generation is reduced to inflection and grammatical ordering of the output.
However, the linguistic markers are assigned somewhat mathematically by rules
selected on the basis of semantic closeness calculations, often resulting in erroneous
structure predictions. To improve the output quality, it would be desirable to have
some form of repair strategies for ill-formed input to the generation component,
restructering the information given in a sensible way.
As it turns out, this task is not easily achieved without being more careful about the
choice of linguistic markers in the transfer component, since the current inconsistencies, irrelevant as they are in the present system, lead to a lot of over-generation
even with simple repair strategies.
In this paper I will present some ideas on methods for repair, and why and how they
will or will not work, given appropriate input from the transfer component. I will
suggest a few changes to the information encoded in the linguistic markers during
transfer, providing more structural information to base the repair strategies on, and
thus avoiding most causes for over-generation.