Bidirectional Linguistic Deduction (BiLD)
Most large natural-language systems with both analysis and generation
capabilities make use of different processing components for the two
directions. However, there are strong arguments for adopting uniform,
bidirectional processing.
In BiLD, a project within the Sonderforschungsbereich 314 "Artificial
Intelligence", the optimal realization of generation and parsing of natural
language within a uniform deduction system is investigated. To this end,
the final results of semantic and graphemic (phonetic) form are encoded in
the same representation languageŠan augmented feature-unification
formalism. The application of methods from theorem proving is investigated.
The most important research results and methods of the project are:
- Application of bidirectional methods to simultaneous feedback
- Reduction of deduction through merging of frequently occurring
sequences of deduction steps
- Depth-first vs. breadth-first control through pruning of the search
space, and implementation of a linguistic formalism with preference values
for best-first search
- Application of tabulated deduction methods (Earley Deduction) for
efficient processing
- A new algorithm for incremental linguistic deduction that can handle
additions to as well as modifications of the input structure
- Different indexing of the same constituents and lexical entries for
interleaved parsing and generation
There is close cooperation with NLP projects at the DFKI and with automated
deduction projects in the Sonderforschungsbereich 314.