Using Entity Information from a Knowledge Base to Improve Relation Extraction

Using Entity Information from a Knowledge Base to Improve Relation Extraction Lan Du1∗, Anish Kumar2, Mark Johnson2 and Massimiliano Ciaramita3 1Faculty of Information Technology, Monash University Clayton, VIC 3800, Australia 2Department of Computing, Macquarie University Sydney, NSW 2109, Australia 3Google Research, Zurich, Switzerland Abstract Relation extraction is the task of ex- tracting predicate-argument relationships between entities from natural language text. This paper investigates whether back- ground information about entities avail- able in knowledge bases such as FreeBase can be used to improve the accuracy of a state-of-the-art relation extraction sys- tem. We describe a simple and effective way of incorporating FreeBase’s notable types into a state-of-the-art relation extrac- tion system (Riedel et al., 2013). Experi- mental results show that our notable type- based system achieves an average 7.5% weighted MAP score improvement. To understand where the notable type infor- mation contributes the most, we perform a series of ablation experiments. Results show that the notable type information im- proves relation extraction more than NER labels alone across a wide range of entity types and relations. 1 Introduction The goal of relation extraction is to extract rela- tional information about entities from a large text collection. For example, given the text “Michael Bay, the director of Transformers, visited Paris yesterday,” a relation extraction system might ex- tract the relationship film director(Michael Bay, Transformers). These tuples can be then used to extend a knowledge base. With the increase in the amount of textual data available on the web, rela- tion extraction has gained wide applications in in- formation extraction from both general newswire texts and specialised document collections such as biomedical texts (Liu et al., 2007). ∗This work was partially done while Lan Du was at Mac- quarie University. A typical relation extraction system functions as a pipeline, first performing named entity recog- nition (NER) and entity disambiguation to link the entity mentions found in sentences to their database entries (e.g., “Michael Bay” and “Trans- formers” would both be linked to their respective database ids). Then the context in which these en- tity mentions co-occur is used to predict the re- lationship between the entities. For example, the path in a syntactic parse between two mentions in a sentence can be used as a feature to predict the relation holding between the two entities. Contin- uing our example, the text pattern feature X-the- director-of-Y (or a corresponding parse subtree fragment) might be used to predict the database relation film director(X,Y). In such a pipeline ar- chitecture, information about the entities from the database is available and can be used to help de- termine the most appropriate relationship between the entities. The goal of this paper is to identify whether that information is useful in a relation ex- traction task, and study such information about the entities with a set of ablation experiments. We hypothesise that information from database entries can play the role of background knowl- edge in human sentence comprehension. There is strong evidence that humans use world knowledge and contextual …

Content truncated to 500 words