DIMACS TR: 99-32

Natural Language Understanding with the Generality Feedback

Author: Boris Galitsky


This study addresses the problem of improving the quality of natural language (NL) understanding by means of the involvement of the additional criterion of correct formal representation of the input inquiry. This criterion is based on the logical compatibility between the input inquiry, translated into the formal language, and the domain, encoded in the same language. Inquiry generality is measured as a normalized number of object tuples, which deliver the translation satisfaction for a given domain. Computational experiments in various domains show that the certain diapason of generality indicates the proper translation. Specific heuristics are developed for the transformation of the translation formula to improve the value of the generality criterion.

Usually, if the generality is too high (there is a large number of the tuples of objects) then the semantic analyzer has likely ignored some syntactic constraints for the translation formula. On the contrary, when the generality is too low (there is no object tuples, satisfying the translation), then the semantic analyzer has likely introduced too many syntactic constraints, and some of them have to be eliminated.

Generality criterion helps in the frequent situations, where the morphological and syntactic processing is insufficient for the construction of the proper inquiry translation into the formal language. Our approach is based on the experimentally verified fact that the user poses inquiries with the reasonable generality. This knowledge helps us to improve the semantic analysis to function in the flexible and expandable domains, where the lexical information is rather limited.

Compatibility feedback by means of generality control poses the specific requirements to the formal language of inquiry translation and domain knowledge representation. The semantic processor is implemented as a logical program with metalanguage support to present the complex semantic rules. Suggested approach allowed us to build the NL understanding system with high inquiry complexity, involving up to 3-4 concepts.

Paper Available at: ftp://dimacs.rutgers.edu/pub/dimacs/TechnicalReports/TechReports/1999/99-32.ps.gz
DIMACS Home Page