[Objective] Aiming at the problems of low negative sampling quality of current translation-based knowledge graph embedding models, which affects the effective embedding of knowledge graph and leads to low representation ability and poor performance, thus a knowledge graph model based on negative sampling in joint relational context was proposed.
[Methods] Firstly, the neighbor of the target instance is extracted from the original knowledge graph and the context vector is generated. Then, according to the properties of adjacent relations that can provide information about the nature or type of a given entity, the Concat aggregation function is used to aggregate the context of the given entity in negative sampling so as to determine the attributes of the entity to be replaced. Finally, the triple embedding of TransE model is combined with the substitution entities with the same attribute to generate negative triples, so as to improve the similarity of the positive and negative triples.
[Results] In the entity linking, the FB15K-237 and WN18RR datasets have improved by 15.3 and 26.7 percentage points respectively compared with the benchmark model. It also saw a 7 percent increase in the related links.
[Limitations] When considering the neighbor relationship, only the semantic information of the relational context is considered, making it hard to determine the relative position, thus the path information needs to be further explored.
[Conclusions] The sampling strategy improves the quality of negative three tuple by improving the similarity between the replaced entity and the entity to be replaced, thus the accuracy of the model gets effectively improved.
李智杰, 王瑞, 李昌华, 张颉. 联合关系上下文负采样的知识图谱嵌入
[J]. 数据分析与知识发现, 10.11925/infotech.2096-3467.2022-0214.
Li Zhijie, Wang Rui, Li Changhua, Zhang Jie. Knowledge Graph Embedding Based on Negative Sampling of Joint Relational Context
. Data Analysis and Knowledge Discovery, 0, (): 1-.