Context and Type Enhanced Representation Learning for Relation Extraction

Relation extraction from plain text aims to extract relational facts between entities in the text, and plays an important role in knowledge graph construction, question answering, and so on. Distant supervision based methods employ an external knowledge graph to automatically generate the training data of relations between entities via linking the entities in the knowledge graph to their mentions in the texts. However, due to the noise in texts, these methods often suffer from the wrong labelling problem. To address this issue, we propose a context and type enhanced representation learning method for relation extraction (CTRL-RE). Specifically, to avoid the noise in texts, the global context information for entities within a given window size in the texts is used to generate the context-based representations of entities. The type of entities is utilized to generate the type-based representations of the entities. Then these two representations of the entities are combined with the representation of relations to form a context and type enhanced knowledge graph representation learning method for relation extraction. Experiments on benchmark datasets show our proposed method can achieve superior performance compared to analogous methods.