Typed Tensor Decomposition of Knowledge Bases for Relation Extraction

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1568–1579, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Typed Tensor Decomposition of Knowledge Bases for Relation Extraction Kai-Wei Chang†∗ Wen-tau Yih♮ Bishan Yang‡∗ Christopher Meek♮ †University of Illinois, Urbana, IL 61801, USA ‡Cornell University, Ithaca, NY 14850, USA ♮Microsoft Research, Redmond, WA 98052, USA Abstract While relation extraction has traditionally been viewed as a task relying solely on textual data, recent work has shown that by taking as input existing facts in the form of entity-relation triples from both knowl- edge bases and textual data, the perfor- mance of relation extraction can be im- proved significantly. Following this new paradigm, we propose a tensor decompo- sition approach for knowledge base em- bedding that is highly scalable, and is es- pecially suitable for relation extraction. By leveraging relational domain knowl- edge about entity type information, our learning algorithm is significantly faster than previous approaches and is better able to discover new relations missing from the database. In addition, when ap- plied to a relation extraction task, our ap- proach alone is comparable to several ex- isting systems, and improves the weighted mean average precision of a state-of-the- art method by 10 points when used as a subcomponent. 1 Introduction Identifying the relationship between entities from free text, relation extraction is a key task for ac- quiring new facts to increase the coverage of a structured knowledge base. Given a pre-defined database schema, traditional relation extraction approaches focus on learning a classifier using tex- tual data alone, such as patterns between the oc- currences of two entities in documents, to deter- mine whether the entities have a particular rela- tion. Other than using the existing known facts to label the text corpora in a distant supervision setting (Bunescu and Mooney, 2007; Mintz et al., ∗Work conducted while interning at Microsoft Research. 2009; Riedel et al., 2010; Ritter et al., 2013), an existing knowledge base is typically not involved in the process of relation extraction. However, this paradigm has started to shift re- cently, as researchers showed that by taking exist- ing facts of a knowledge base as an integral part of relation extraction, the model can leverage richer information and thus yields better performance. For instance, Riedel et al. (2013) borrowed the idea of collective filtering and constructed a ma- trix where each row is a pair of entities and each column is a particular relation. For a true entity- relation triple (e1, r, e2), either from the text cor- pus or from the knowledge base, the correspond- ing entry in the matrix is 1. A previously unknown fact (i.e., triple) can be discovered through ma- trix decomposition. This approach can be viewed as creating vector representations of each relation and candidate pair of entities. Because each entity does not have its own representation, relationships of any unpaired entities cannot be discovered. Al- ternatively, Weston et al. (2013) created two types of embedding – one …

Content truncated to 500 words