基于 MpBERT-BiGRU 的中文知识图谱补全-《计算机技术与发展》

文章信息/Info

作者:: 田昊¹; 2 ; 张骁雄¹ ; 刘文杰² ; 刘浏¹; 3 ; 刘姗姗¹ ; 丁鲲¹; 1. 国防科技大学第六十三研究所,江苏南京 210007;
2. 南京信息工程大学计算机与软件学院,江苏南京 210044;
3. 宿迁学院,江苏宿迁 223800

Author(s):: TIAN Hao1; 2 ; ZHANG Xiao-xiong1 ; LIU Wen-jie2 ; LIU Liu1; 3 ; LIU Shan-shan1 ; DING Kun1; 1. The Sixty-third Research Institute,National University of Defense Technology,Nanjing 210007,China;
2. School of Computer and Software,Nanjing University of Information Science and Technology,Nanjing 210044,China;
3. Suqian University,Suqian 223800,Chin

Keywords:: knowledge graph completion; pooling strategy; BiGRU; BERT; linking prediction

摘要:: 知识补全是提高知识图谱质量的关键技术,为了更好地利用中文知识图谱,该文对中文知识图谱补全进行研究。针对大多数研究聚焦于英文数据集,缺少中文知识补全数据集的情况,在已有数据集的基础上,该文构建了中文 UMLS+ownthink 数据集。现有知识图谱补全方法大多忽视 BERT 模型表征能力不足、位置信息学习能力弱的问题,且未考虑中文文本特征复杂、语序依赖性强的特点,因此提出一种名为 MpBERT-BiGRU 的中文知识图谱补全模型,利用平均池化策略有效缓解 BERT 模型表征能力弱的问题,并通过 BiGRU 网络强化特征信息,提高位置信息学习能力;同时将三元组转化为文本序列,结合实体描述信息作为模型的输入,利用背景知识丰富实体信息。链接预测实验结果表明,该方法在平均排名(Mean Rank,MR) 指标上相比传统方法提高 10. 39,前 10 命中率( Hit@ 10) 指标提高 4. 63% ,验证了模型在中文语料库上的有效性。

Abstract:: Knowledge graph completion is an important technology to improve knowledge graph quality. To make better use of Chineseknowledge graph,Chinese knowledge 　graph completion was studied. In view of the fact that most studies focus on English data sets andlack Chinese knowledge completion data sets,we construct Chinese UMLS+ownthink data sets on the basis of existing data sets. Most ofthe existing knowledge graph completion methods ignored insufficient sequence representation ability and weak position learning ability ofBERT model,and didn’t take into account the complex textual feature and strong word order dependence of Chinese text. Therefore, MpBERT-BiGRU model was proposed for Chinese knowledge graph completion. Mean-pooling strategy was used to improve the sequencerepresentation ability of BERT model,and BiGRU was adopted to strengthen the feature information and improve the position learningability. Meantime,the triples were transformed 　into text sequence and fed to input layer. Combined with description information,the　entities＇ information were enriched by abundant background knowledge. The experimental results of linking prediction showed that theproposed method can improve the Mean Rank ( MR) index by 10. 39 and the top 10 hit rate ( Hit@ 10) index 　by 4. 63% compared withtraditional methods,verifying the effectiveness of this model on the Chinese data set.