From Logical Forms to Knowledge Frames: An Experiment on Scientific Text

Although logical forms adequately represent the local information contained in sentences, they are not a malleable, robust representation for reasoning and must be converted into such a representation. This research attempts to determine whether it is possible to accomplish such a conversion in a straightforward way into a knowledge base which can be reasoned about. A method is described and implemented for converting the logical forms into a network of knowledge frames. A modular approach is taken, using the standalone parser of the Alvey Tools to produce the logical forms. An ontology of concepts is built to contextualize the interpretation of the logical forms and to provide a well-formed model of world knowledge. The ontology plus the structure of the logical forms provide most of the information for the conversion. Constraint satisfaction and phrasal information are used to guarantee the well-formedness of the knowledge frames with respect to the ontology. Thus the converter could act as a conceptual filter on the output of the parser by reducing the number of interpretations kept from the parser's output, though this has not been demonstrated. Both the design and implementation of the converter are discussed. This research raised several issues, such as the large amount of hand work needed to construct the ontology, problems with knowledge representation for quantification, and structural and lexical ambiguities which led to a huge number of parses for some sentences (which in some cases required so much virtual memory that parsing was not completed). The large number of ambiguities in the parser's output are a major weakness in the modular approach. 1This work was supported in part by the Division of Instrumentation and Resources of the NSF (grant number DIR-88-14522). Submitted in partial fulfillment of the requirements for the degree of M.S. in Computer Science.