Graph Attention Network Entity Relation Extraction Model With Denoising Gate

Entity relation extraction plays a vital role in natural language processing.It aims to identify the semantic relation between entity pairs from plain text,and has been widely used in knowledge base construction and question answering systems.As a research hotspot in this field,distant supervision relation extraction method combines the relation instance in the knowledge base as auxiliary information,and labels the data automatically by aligning it with unstructured text.Although distant supervision method can effectively alleviate the problem of insufficient training corpus,it also introduces data noise due to incorrect labeling.Most of the existing work adopts a multi-instance learning method to reduce the impact of data noise on relation extraction,which treats all the sentences that mention the same entity pair as a bag,and applys the attention mechanism to select the important sentences in the bag.However,when the data in the bag is too sparse(for example,80% of the bags in the NYT dataset have only one sentence),the attention mechanism may be low-effective.In addition,although the existing method based on dependency trees can effectively capture long-range dependency between words,they ignore the difference in the importance of different words for expressing relation.As a result,these methods cannot selectively focus on the most relevant syntactic structure for relation extraction in the dependency tree,which makes them difficult to extract the semantic relation between entities accurately and effectively.In order to solve the above problems,this paper proposes a Graph Attention Network with Denoising Gate(GAT?DGATE)entity relation extraction model.Specifically,the model transforms the dependency tree into a weighted directed graph with graph attention network.It selects sub-structures that are more critical to the relation expression through node relevance,and effectively eliminates the interference of redundant information in the sentence.Therefore,it expresses the semantic relation between entities more accurately.Furthermore,this paper also proposes a novel denoising gate mechanism,which calculates the gate value according to the similarity between each sentence and the label of the bag,and applys the average pooling operation of the gate value instead of the normalization operation of attention weight.It avoids the situation where the attention mechanism is difficult to take effect when there is only one sentence in the bag,which further enhances the robustness of the model.In addition,in consideration of the fact that distant supervision method usually generates more negative data than positive data,this paper also adopts s Gradient-based One-Side Sampling to solve the problem of class imbalance.This paper designs and conducts comparative experiments based on the NYT dataset which is widely used in the task of distant supervision relation extraction.The experimental results show that compared with the mainstream methods in the industry,the GAT?DGATE model proposed in this paper performs better on the evaluation indicators such as PR curve and AUC,Precision@N,etc.,which verifies the effectiveness of the algorithm.Moreover,this paper also makes a qualitative evaluation of the power of graph attention network and denoising gate through visual analysis and case analysis.The results show that the proposed method has good interpretability.