Abstraction of the EDR Concept Classification and its Effectiveness in Word Sense Disambiguation

Abstract

The relation between the degree of abstraction of a concept and the explanation capability (validity and coverage) of conceptual description which is the constraint held between concepts is clarified experimentally by performing the operation called concept abstraction. This is the procedure that chooses a set certain of lower level concepts in a concept hierarchy and maps the set to one or more upper level (abstract) concepts. We took the three abstraction techniques of a flat depth, a flat size, and a flat probability method for the degree of abstraction. By taking these methods and degrees as a parameter, we applied the concept abstraction to the EDR Concept Classifications and performed word sense disambiguation test. The test set and the disambiguation knowledge were extracted as a co-occurrence expression from the EDR Corpora. Through the test, we found that the flat probability method gives the best result. We also carried out an evaluation by comparing the abstracted hierarchy with that of human introspection and found the flat size method gives the most similar results to human. These results would contribute to clarify the appropriate detailed-ness of a concept when given an application purpose of a concept hierarchy.

Resources

Details