基于知识表示学习的KBQA答案推理重排序算法

KBQA answer inference re-ranking algorithm based on knowledge representation learning

现有的知识库问答(KBQA)研究通常依赖于完善的知识库,忽视了实际应用中知识图谱稀疏性这一关键问题。为了弥补该不足,引入了知识表示学习方法,将知识库转换为低维向量,有效摆脱了传统模型中对子图搜索空间的依赖,并实现了对隐式关系的推理,这是以往研究所未涉及到的。其次,针对传统KBQA在信息检索中常见的问句语义理解错误对下游问答推理的错误传播,引入了一种基于知识表示学习的答案推理重排序机制。该机制使用伪孪生网络分别对知识三元组和问句进行表征,并融合上游任务核心实体关注度评估阶段的特征,以实现对答案推理结果三元组的有效重排序。最后,为了验证所提算法的有效性,在中国移动RPA知识图谱问答系统与英文开源数据集下分别进行了对比实验。实验结果显示,相比现有的同类模型,该算法在hits@n、准确率、F_(1)值等多个关键评估指标上均表现更佳,证明了基于知识表示学习的KBQA答案推理重排序算法在处理稀疏知识图谱的隐式关系推理和KBQA答案推理方面的优越性。

Existing research on knowledge base question answering(KBQA)typically relies on comprehensive knowledge bases,but often overlooks the critical issue of knowledge graph sparsity in practical applications.To address this shortfall,this paper introduced a knowledge representation learning method that transforms knowledge bases into low-dimensional vectors.This transformation effectively eliminated the dependence on subgraph search spaces inherent in traditional models and achieved inference of implicit relationships,which previous research had not explored.Furthermore,to counter the propagation of errors in downstream question-answering inference caused by semantic understanding errors of questions in traditional KBQA information retrieval,this paper introduced an answer inference re-ranking mechanism based on knowledge representation learning.This mechanism utilized pseudo-twin networks to represent knowledge triplets and questions separately,and integrated features from the core entity attention evaluation stage of upstream tasks to effectively re-rank the answer inference result triplets.Finally,to validate the effectiveness of the proposed algorithm,this paper conducted comparative experiments on the China Mobile RPA knowledge graph question-answering system and an English open-source dataset.Experimental results demonstrate that,compared to existing models in the same field,the proposed method performs better in multiple key evaluation indicators such as hits@n,accuracy,and F_(1)-scores,proving the superiority of the proposed KBQA answer inference re-ranking algorithm based on knowledge representation learning in handling implicit relationship inference in sparse knowledge graphs and KBQA answer inference.