Enhancing Joint Entity and Relation Extraction with Language Modeling and Hierarchical Attention

Abstract

Both entity recognition and relation extraction can benefit from being performed jointly, allowing them to enhance each other. However, existing methods suffer from the sparsity of relevant labels and strongly rely on external natural language processing tools, leading to error propagation. To tackle these problems, we propose an end-to-end joint framework for entity recognition and relation extraction with an auxiliary training objective on language modeling, i.e., learning to predict surrounding words for each word in sentences. Furthermore, we incorporate hierarchical multi-head attention mechanisms into the joint extraction model to capture vital semantic information from the available texts. Experiments show that the proposed approach consistently achieves significant improvements on joint extraction task of entities and relations as compared with strong baselines.