【摘要】 关系抽取是提取实体间关系的一项重要的自然语言处理任务。最近的研究发现，预训练BERT模型在自然语言处理任务中取得了非常好的效果。此后，诞生了大量使用预训练BERT模型处理关系抽取任务的方法，其中具有代表性的是R-BERT方法。但是，该方法在实现时未考虑主语实体与宾语实体在语义上的差异，以及全局语义信息对关系抽取任务准确性的影响。通过设置两个不同的全连接层来分别提取主语实体和宾语实体的信息，从而将主语实体与宾语实体在语义上的差异引入模型的学习过程中。此外，还在原有的信息融合模块后面添加了一层带有激活函数的新全连接层来将高维全局语义信息与实体对充分融合。将融合了语义差异与全局语义信息的R-BERT简称为GR-BERT。通过在中文人物关系抽取数据集上进行实验，结果表明新提出的GR-BERT的效果较原始R-BERT取得了显著提升，从而验证了新方法GR-BERT的有效性。更多还原

【Abstract】 Relationship extraction is an important natural language processing task that involves extracting relationships between entities. Recent research has shown that the pre-trained BERT model achieves very good results in natural language processing tasks. Since then, many methods using pre-trained BERT model for relationship extraction have been developed, with R-BERT being a representative example. However, this method does not consider the semantic differences between the subject and object entities or the impact of global semantic information on the accuracy of relationship extraction tasks. This paper addresses this by setting up two separate fully connected layers to extract information for the subject and object entities,thereby incorporating the semantic differences between them into the model’s learning process.Additionally, a new fully connected layer with an activation function is added after the existing information fusion module to fully integrate high-dimensional global semantic information with entity pairs. The proposed model, which integrates semantic differences and global semantic information, is referred to as GR-BERT. Experiments on a Chinese entity relationship extraction dataset show that GR-BERT significantly improves the original R-BERT, thereby validating its effectiveness.更多还原