基于语义联合的知识库问答方法

Knowledge Base Question Answering Method Based on Semantic Fusion

摘要: 基于知识图谱的问答(KBQA)是问答系统的重要组成部分。然而，现有多数知识图谱问答系统侧重于回答单个三元组查询的简单问题，对回答涉及多个实体和关系的复杂问题的正确率较低。为了提高正确率，本文采用对比学习方法来计算语义相似度，并提出了语义联合模型框架，将实体消歧和关系匹配任务进行联合建模，以避免误差传递问题。最终使用本文的方法在CCKS2019KBQA数据集上进行实验，实验结果表明，与BERT模型相比，对比学习模型在计算语义相似度方面更具优势，并且语义联合建模的效果也优于先进行实体消歧再进行关系匹配的方法。

Abstract: Translation: Knowledge-based question answering (KBQA) is an important component of question answering systems. However, most existing KBQA systems primarily focus on answering simple questions that involve single triple pattern queries, resulting in lower accuracy when it comes to answering complex questions involving multiple entities and relationships. To improve the accuracy, this paper adopts contrastive learning to calculate semantic similarity and proposes a semantic joint modeling framework that combines entity disambiguation and relationship matching tasks to avoid error propagation. The proposed method is evaluated on the CCKS2019KBQA dataset, and the experimental results demonstrate that compared to the BERT model, the contrastive learning model has a greater advantage in computing semantic similarity, and the effectiveness of semantic joint modeling is superior to the approach of first performing entity disambiguation and then relationship matching.