Spotting Knowledge Base Facts in Web Texts

Knowledge bases (KB) such as DBpedia, YAGO and Freebase have been constructed by harvesting facts from high-quality data sources and incorporating community contributions. Accurately detecting occurrences of these KB facts in complementary sources (sources other than where they were extracted from) is crucial for fact validity assessments and deriving occurrence statistics. In this paper we consider fact spotting – the task of automatically discovering the mentions of KB facts in text documents. Our fact spotting methodology follows a two-stage approach. First, we perform similarity-based matching of noun phrases with labels of KB entities and dependency path structures with patterns of KB relations. Next, we perform joint matching of mentions, entities, paths, relations, and their textual locations by encoding them into variables of an integer linear program. We evaluate our method by spotting Freebase facts in biographies on the Web.