Abstract: The zero-shot knowledge graph relation prediction task aims to infer fact triples containing unseen relations based on existing fact triples in the knowledge graph. Most existing methods fail to fully utilize the descriptive information of entities in the knowledge graph, focusing only on textual information of relations and graph structural information. Moreover, the textual information of relations relies solely on the dataset's inherent descriptions, making it difficult to capture the detailed semantics of relations and the overall semantic information of the knowledge graph triples. To address these issues, a dual-channel embedding fusion-based method for zero-shot knowledge graph relation prediction is proposed. First, a structural encoder and a semantic encoder are employed to generate the structural and semantic features of the entity components in the triples, respectively. Subsequently, a dual-channel embedding fusion module is used to integrate the structural and semantic features of the entities in the knowledge graph. Meanwhile, the text generation capability of large language models is leveraged to produce enhanced textual descriptions for relations with enriched semantics. During training, the semantically enhanced features of relations and the multimodal features of entities are aligned in the vector space to ensure semantic consistency. Finally, for zero-shot relations, their semantically enriched textual descriptions are utilized to accomplish the zero-shot knowledge graph relation prediction. Experimental results demonstrate improvements across various metrics in the zero-shot knowledge graph completion task, validating the effectiveness of the proposed model.

Key words: knowledge graph completion (KGC), zero-shot KGC, link prediction, multi-modal fusion, generative adversarial networks

摘要： 零样本知识图谱关系预测任务旨在根据知识图谱已有的事实三元组来推理预测包含未见关系的事实三元组。已有的方法大多未充分利用知识图谱实体的描述信息，仅考虑了关系的文本信息和图结构信息；针对关系的文本信息仅依赖于数据集本身的文本描述，难以捕捉关系的详细语义与知识图谱三元组整体的语义信息。为解决以上问题，提出一种基于双通道嵌入融合的零样本知识图谱关系预测方法，首先，使用结构编码器和语义编码器分别生成三元组实体部分的结构特征与语义特征。随后，通过双通道嵌入融合模块融合知识图谱实体的结构特征和语义特征，同时利用大语言模型的文本生成能力，生成关系文本语义增强后的描述信息。经过训练使关系增强后的语义特征与实体的多模态特征在向量空间中满足语义一致性。最后，针对零样本关系，根据其语义丰富后的文本描述信息即可完成零样本知识图谱关系预测。实验结果显示在零样本知识图谱补全任务的各项指标上都取得了提升，验证了所提模型的有效性。

关键词: 知识图谱补全, 零样本知识图谱补全, 关系预测, 多模态融合, 生成对抗网络