Abstract Reidsma
This paper reports on the development of a system that constructs representations of the meaning of sentences using rules of grammar and a dictionary of word meanings. Furthermore an algorithm will be presented that can be used for ambiguity resolution. This algorithm evaluates the meaning representation of a sentence on the amount of complementary or contradictory information it contains.The meaning of words and sentences is expressed using knowledge graphs, a semantic network formalism, which has been extended with link weights to support the algorithms described in this paper and to improve the expressivity of the formalism.
Up till now most approaches to the creation of the so-called sentence graphs (using knowledge graphs or other formalisms) have been dependend on manual control of this unification process. The unification algorithm presented in this paper creates sentence graphs automatically, using knowledge about the grammatical function of parts of word graphs. The algorithm completes the unification of word graphs into sentence graphs by removing redundant information using a similarity measure expressing the amount of contradictory or complementary information introduced into the sentence graphs through the different word graphs and their grammatical relations.
The unification algorithm may result in several possible sentence graphs, for example in cases of lexical or syntactical ambiguity. The second algorithm presented in this paper compares these sentence graphs and assigns evaluation values expressing the relative quality of each the sentence graphs with respect to the other possible graphs for the same sentence. For this a variant of the same similarity measure is used.
The resulting system has been tested on some cases of PP-attachment desambiguation and lexical desambiguation and seems to perform well.
Last modified $Date: 2001/10/04 13:39:48 $ by Parlevink Webmaster