Generating and Evaluating Triples for Modelling a Virtual Environment

Our purpose is to extract RDF-style triples from text corpora in an unsupervised way and use them as preprocessed material for the construction of ontologies from scratch. We have worked on a corpus taken from Internet websites and describing the megalithic ruin of Stonehenge. Using a shallow parser, we select functional relations, such as the syntactic structure subject-verb-object. The selection is done using prepositional structures and frequency measures in order to select the most relevant triples. Therefore, the paper stresses the choice of patterns and the filtering carried out in order to discard automatically all irrelevant structures. At the same occasion, we are experimenting with a method to objectively evaluate the material generated automatically.