SETFF: A Semantic Enhanced Table Filling Framework for Joint Entity and Relation Extraction

Hao Li, Md Tauhidul Islam, Kehan Huangliang, Zugang Chen, Kaiyang Zhao, Haixian Zhang

Abstract: In the study of text understanding and knowledge graph construction, the process of extracting entities and relations from unstructured text is crucial. Lately, joint extraction has achieved more significance in this context. Among them, table filling based method has attracted a lot of research in solving the problem of overlapping relation in complex scenarios. However, most existing table filling works need to deal with many invalid and redundant filling processes. At the same time, some semantic information is not fully considered. For instance, a token should have differentiated semantic representation when decoding triples under different relations. Moreover, the global association information between different relations is not fully utilized. In this paper, we propose a joint extraction framework: SETFF, based on table filling. Firstly, the proposed method filters out the possible relations in sentences through a relation filtering module. Then following the attention mechanism, the pre-trained relation embeddings are used to enhance the differential representation of token semantics under specified relations and obtain the mutual prompting information between different relations. In addition to these, extensive experimental results show that SETFF can effectively deal with the overlapping triples problem and achieve significant performance on two public datasets.