基于位置编码与实体交互信息的关系抽取方法

关系抽取作为信息抽取领域的重要研究课题, 其主要目的是抽取句子中已标记实体对之间的语义关系, 对句子语义理解及知识库构建有着重要作用. 针对现有抽取方法中未能充分利用单词位置信息和实体间的交互信息导致重要特征丢失的问题, 本工作提出一种基于位置编码与实体交互信息的关系抽取方法(BPI-BERT). 首先将新型位置编码融入BERT预训练语言模型生成的词向量中后使用平均池化技术得到实体和句子向量, 再利用哈达玛乘积构造实体交互信息, 最后将实体向量、句子向量及交互信息向量拼接得到关系向量并输入到Softmax分类器进行关系分类. 实验结果表明BPI-BERT在精准率和 F1上较现有方法有提高, 证明了BPI-BERT的有效性.

Relation Extraction Method Based on Entity Interaction and Position Encoding

As an important research program in the field of information extraction, relation extraction is mainly aimed at extracting semantic relations between labeled entity pairs in sentences, which plays an important role in sentence semantic understanding and knowledge base construction. The existing extraction methods fail to make full use of the word position information and the interaction information between entities, which leads to the loss of effective features in relation extraction. To solve this problem, this study proposes a relation extraction method BPI-BERT based on the interaction information between position encoding and entities. The novel position coding is integrated into the word vector generated by the BERT pre-trained language model, and the entity and sentence vectors are obtained through the average pooling technology. The Hadamard product is used to construct the entity interaction information. Finally, the entity vector, sentence vector, and interaction information vector are stitched together to obtain the relation vector which is then input to the Softmax classifier for relation classification. The experimental results show that the precision and F1 of BPI-BERT are significantly improved compared with those of the existing methods, and thus the effectiveness of BPI-BERT is proved.