Bootstrapping Joint Entity and Relation Extraction with Reinforcement Learning

Extracting entities and relations for types of interest from text is important for knowledge graph construction. Previous methods of entity and relation extraction rely on human-annotated corpora and adopt an incremental pipeline, which require abundant human expertise and are vulnerable to errors cascading. In this paper, we present ROTATE, a novel approach for jointly bootstrapping entity and relation extraction with reinforcement learning. The bootstrapping process of ROTATE consists of three bootstrapping sub-processes, which extract head entity, tail entity, and relation, respectively. Each sub-process starts with a few seed instances, then generates patterns and expands the seed set to start the next iteration. In particular, we propose a joint pattern scoring strategy in which the scores of patterns in each bootstrapping sub-process also take the seed information of the other two sub-processes into account. Moreover, we introduce reinforcement learning to solve the semantic drift problem in the bootstrapping process by formulating the seed expansion problem as a sequential decision making problem, and design a reward function that considers both seed quality and quantity. Experimental results on a collection of sentences from news articles confirm the effectiveness of our approach.