Abstract:

Aiming at the problems of incomplete semantic representation of hyper-relational facts as well as insufficient task accuracy and generalization ability caused by the qualifier sparsity in Hyper-relational Knowledge Graphs （HKGs）， a Large language model-driven method for Qualifier Enhancement in HKGs （L-EQs） was proposed. First， semantically rich description information was obtained from the associated knowledge bases to mitigate semantic ambiguity caused by the same labels of entities and relations. Second， prompt templates were used to guide the LLM in generating the matched qualifiers and their semantic explanations as external knowledge. To ensure the quality of the qualifiers， an iterative prompting strategy was employed to filter the generated qualifiers， with multiple iterations conducted until the predetermined quantity was reached. Third， embeddings of external knowledge were constructed through the semantic explanations， and an external knowledge aggregation module was introduced to suppress noise， thereby providing high-quality semantic support for the downstream task by integrating external knowledge and original knowledge embeddings. Finally， the task results were predicted through a decoder， and extensive experiments were conducted on the open-source datasets WikiPeople and WD50K to validate the effectiveness of the L-EQs. Experimental results show that compared to the optimal results of the baseline models， the Mean Reciprocal Rank （MRR）， Hit@1， Hit@5， and Hit@10 of the L-EQs are increased by 8.4， 4.7， 5.3， and 6.8 percentage points， respectively， on the WikiPeople dataset， and are improved by 4.8， 2.3， 2.7， and 3.0 percentage points， respectively， on the WD50K. L-EQs addresses the qualifier sparsity in HKG and the issues caused by the incomplete semantic representations.

Key words: Hyper-relational knowledge graph (HKG), Large Language Model (LLM), Hyper-relational Knowledge Graph Completion (HKGC), hyper-relational fact, qualifier

摘要：

针对超关系知识图谱（HKG）中限定符稀疏现象导致的超关系事实语义表征不完整以及任务精度和泛化能力不足的问题，提出基于大语言模型（LLM）的HKG限定符增强方法（L-EQs）。首先，通过关联知识库获取语义丰富的描述信息，以减轻实体和关系因相同标签而导致的语义混淆；其次，使用提示词模板引导LLM生成匹配的限定符及其语义解释作为外源知识，而为了保证限定符的质量，采用迭代式提示策略对生成的限定符进行筛选，并通过多次迭代生成限定符直至达到预定数量；再次，通过语义解释构建外源知识嵌入，并利用外源知识聚合模块抑制噪声，从而将外源知识与原始知识嵌入融合，为下游任务提供高质量的语义支持；最后，通过解码器预测任务结果，并在开源数据集WikiPeople和WD50K上进行广泛实验，以验证L-EQs的有效性。实验结果表明，L-EQs的平均倒数排名（MRR）、Hit@1、Hit@5和Hit@10相较于基线中的最优结果在WikiPeople数据集上分别提升了8.4、4.7、5.3和6.8个百分点，在WD50K数据集上分别提升了4.8、2.3、2.7和3.0个百分点。L-EQs解决了HKG中限定符稀疏现象和超关系事实语义表征不完整带来的问题。

关键词: 超关系知识图谱, 大语言模型, 超关系知识图谱补全, 超关系事实, 限定符