A relation aware embedding mechanism for relation extraction

Extracting possible relational triples from natural language text is a fundamental task of information extraction, which has attracted extensive attention. The embedding mechanism has a significant impact on the performance of relation extraction models, and the embedding vectors should contain rich semantic information that has close relevance to the relation extraction task. Driven by this motivation, we propose a R elation A ware Embedding Mechanism (RA) for relation extraction. In specific, this mechanism incorporates the relation label information into sentence embedding by leveraging the attention mechanism to distinguish the importance of different relation labels to each word of a sentence. We apply the proposed method to three state-of-the-art relation extraction models: CasRel, SMHSA and ETL-Span, and implement the corresponding models named RA-CasRel, RA-SMHSA and RA-ETL-Span. To evaluate the effectiveness of our method, we conduct extensive experiments on two widely-used open datasets: NYT and WebNLG, and compare RA-CasRel, RA-SMHSA and RA-ETL-Span with 12 state-of-the-art models. The experimental results show that our method can effectively improve the performance of relation extraction. For instance, RA-CasRel reaches 91.7% and 92.4% of F1-score on NYT and WebNLG, respectively, which is the best performance among all the compared models. We have open sourced the code of our proposed method in [ 1 ] to facilitate future research in relation extraction.