Universal Schema for Knowledge Representation from Text and Structured Data

Опубликовано 27 июня 2016, 18:37

Entities and the relations among them are central for representing our knowledge about the world. My work concentrates on discovering relations from information sources available to us, including unstructured text corpora and structured data. Previous work for relation extraction can be classified into two categories in terms of how they represent relations. The first is based on supervised learning, representing relations using pre-defined types from knowledge bases. This approach relies on human efforts to define relation types. The second generalizes the Open IE style relation extraction, representing relations as clusters of textual patterns. This assumes semantic equivalence among patterns falling into the same cluster, failing to represent the diversity and ambiguity of the patterns. I will present a new approach, Universal Schema – the union of all relations seen among surface patterns and available structured knowledge bases. This representation preserves the diversity and ambiguity of textual patterns and allows us to generalize among them. In this talk, I will explain how to perform relation extraction in universal schema. We formalize the task as matrix completion and employ matrix factorization to learn implications among relations. Experiments demonstrate that using universal schema for relation extraction provides new state-of-the-art accuracy. We also extend universal schema to entity type extraction.