Navigating Extracted Data with Schema Discovery

Open Information Extraction (OIE) is a recently-introduced type of information extraction that extracts small individ- ual pieces of data from input text without any domain- specific guidance such as special training data or extrac- tion rules. For example, an OIE system might discover the triple Frenzy, year, 1972 from a set of documents about movies. Because OIE is domain-independent, it promises to help users when they have a corpus of structured data, but that structure is unknown, such as when browsing a novel domain or formulating a query. We can describe the struc- ture to the user by displaying a relational schema that fits the extracted data. Unfortunately, the extractions do not carry full schema information: we have extracted values, but not the cor- rect relations, their rows, or their columns. In response we propose TGen, an algorithm for schema discovery, which automatically derives a high-quality relational schema for the extracted data. Dierent applications have dierent schema-design requirements, which can be encoded as input to TGen. We show that our data-mining approach runs in minutes on millions of documents while still resulting in schemas that are useful for exploring unfamiliar data or for composing queries over extracted data.