【摘要】 基于知识图谱嵌入模型,提出一种知识图谱嵌入评分与链路评分相结合的评分方法,以解决中文领域的多跳知识图谱问答任务,与传统的单跳知识问答方法相比适用性更广.该方法在搜索最优答案的同时构建一个查询链路,通过查询给出答案集合,从而有效缓解了现有方法中遗漏答案的情况.在NLPCC-MH数据集上的实验结果表明,该方法在多跳问题上的平均F₁值为0.653,显著优于对比方法.真实知识图谱通常存在链路缺失的情况,实验以随机丢弃25%三元组的方式模拟了知识图谱的稀疏性,结果表明该方法在这种情况下仍然有效.更多还原

【Abstract】 Based on the knowledge graph embedding model, we proposed a scoring method combining knowledge graph embedding scoring and link scoring to solve multi-hop knowledge graph question answering task in the Chinese domain, which had wider applicability compared with the traditional single-hop knowledge question answering methods. The method constructed a query link while searching for the optimal answer, and gave the answer set by query, which effectively alleviated the situation of missing answers in existing methods. The experimental results on the NLPCC-MH dataset show that the average F₁ value of the method on multi-hop problems is 0.653, which is significantly better than the comparison method. Real knowledge graphs usually have missing links, and the experiments simulate the sparsity of knowledge graphs by randomly discarding 25% triples, the results show that the method is still effective in this case.更多还原