Recently, most relational extraction models usually mitigate the adverse effects of noise in sentences for the prediction results, utilizing different tools of natural language processing that to capture high-level features in sentences combined. However, these attention mechanisms do not manage to exploit as much as possible the semantic information of certain keywords that have relational expressive information in the sentence. Therefore, this paper proposes a model based on the keyword's

more »

... ntion mechanism, which is a novel attention mechanism based on the keywords of relational expression related. In particular, the proposed attention mechanism utilizes a linear-chain conditional random field that combines entity-pair features, similarity features between entity-pair features, and its hidden vectors to compute each word's marginal distribution defined as the attention weight. Experimental results show that the method can focus on keywords with relational expression semantics in sentences without using sophisticated tools and achieves performance improvements on the SemEval-2010 Task 8 dataset.