Three Experiments on Mining the Web for Ontology and Lexicon Learning

This paper describes an approach to alleviating the well-known problem of the knowledge acquisition bottleneck in knowledgebased systems. In knowledge-based, meaning-oriented natural language processing, the core knowledge resources are a semantic lexicon and an ontological world model in terms of which lexical meaning is expressed. We describe a mutual bootstrapping approach whereby existing resources are used to create additional resources that are then added to the original resources. Thus, text understanding bootstraps the learning process , which in turn boosts the knowledge resources underlying text understanding. Specifically, in our experiments an existing ontology and an existing semantic lexicon are used by the ontologicalsemantic text analyzer OntoSem to analyze sentences mined from the web and containing specific words unknown to the system to generate candidates for ontological concepts and lexicon entries that at the time are not part of OntoSem’s static knowledge resources. The experiments described in the paper have as their goal a) empirical determination of the number of senses for a word; b) automatic creation of ontological concepts (named sets of property-value pairs) describing the meanings of word senses; and c) suggesting the location of the newly acquired concepts in the ontological network. The experimental environment described is also used for empirical validation of ontological property values in concepts originally encoded by knowledge engineers.