Determining the Best Combination of Word Senses.

The early versions of the uK analyzer at this point simply tried all of the possible combinations of word senses. Each combination activates the applicable constraints, which are combined into a total score for the combination. The combination with the best total score is chosen as the basic Semantic Dependency Analysis, the core TMRs to which other microtheories (such as aspect and coreference) can be applied. In the example sentence, the following choices were made:

1. ``a-traves-de'' is INSTRUMENT, since its LOCATION meaning would require ``adquirir'' to be a PHYSICAL-OBJECT.
2. ``en'' is LOCATION, since its TEMPORAL meaning requires ``espana'' to be a TEMPORAL-OBJECT.
3. ``adquirir'' maps into ACQUIRE, since its LEARN sense requires ``Dr-Andrew'' to be INFORMATION.
4. ``Dr-Andrew'' is an ORGANIZATION, since its HUMAN meaning cannot be the THEME of an ACQUIRE concept.
5. uK currently has trouble choosing between the CORPORATION and SOCIAL-EVENT meaning of ``compania,'' the object of the ``a-traves-de'' PP. Both can have locations in Spain, and both can be INSTRUMENTS of EVENTs. At this point, uK needs to add information into the ontology that ORGANIZATIONS can typically fill the INSTRUMENT slot of ACQUIRE acts, but SOCIAL-EVENTS cannot.

It must be stressed that all of these choices resulted from the fact that the combination of all scores from all the semantic constraints for this combination of word senses was judged superior to any other combination of word senses.

The Mikrokosmos Project at NMSU is one of the first, large-scale attempts at a knowledge-based machine translation system. We have successfully implemented the first and central stage of Semantic-Dependency-Structure building. This involved the creation of a large, language independent ontology which interacts with the Spanish semantic lexicon. The uK analyzer extracts semantic constraints from these two sources, analyzes them using a sophisticated graph search function, and determines the combination of choices that leads to the best overall score.

Previous to the work on Hunter-Gatherer, the semantic analyzer performed its search through an exhaustive listing of combinations of word senses. This approach worked for many inputs, but the process was tedious, and, for some longer sentences, the processing could consume several days. For two or three sentences, no results could be obtained at all.