Toward Ontology-based Knowledge Extraction from Web Data with the Lexicalization of Ontology for Korean QA System

Most of knowledge is written in natural language and structured knowledge base includes the partially limited information of them. In QA system perspective, the quality of knowledge base is depends on how it covers the knowledge to answer user’s questions. To deal with this knowledge base construction problem, we define the natural language question sets and answer documents which contains knowledge enough to answer. And we manually generate triples in DBpedia schema to use as a test dataset. In this paper we present a framework for automatically structuring of knowledge from unstructured data and semi-structured data with pattern learning method to cover the knowledge to answer natural language question dataset. Our platform extracts entities and its relations in triple structure, and its schema is based on ontological knowledge representation. To achieve this goal, our pattern learning method is based on lexicalization of ontology. For unstructured data, we lexicalize ontological representation from natural language from its dependency structure. And for semi-structured data, the headers in table are also lexicalized. We have instantiated the system for Korean. To evaluate our system, we build gold standard triple datasets from natural language, and use F-measure and coverage measurement.