%A Yu Chuanming, Zhang Zhengang, Kong Lingge %T Comparing Knowledge Graph Representation Models for Link Prediction %0 Journal Article %D 2021 %J Data Analysis and Knowledge Discovery %R 10.11925/infotech.2096-3467.2021.0491 %P 29-44 %V 5 %N 11 %U {https://manu44.magtech.com.cn/Jwk_infotech_wk3/CN/abstract/article_5219.shtml} %8 2021-11-25 %X

[Objective] This study systematically reviews the internal mechanism and influencing factors of knowledge graph representation models, aiming to investigate their impacts on specific tasks. [Methods] For the link prediction task, we compared the performance of translation-based and semantic matching-based knowledge graph representation models on FB15K, WN18, FB15K-237 and WN18RR datasets. [Results] With the Hits@1 indicator, the TuckER model generated the best value on WN18, FB15K-237 and WN18RR datasets (0.946 0, 0.263 3 and 0.443 0, respectively), while the ComplEx model yielded the highest value on FB15K dataset (0.731 4). [Limitations] We only compared the effects of knowledge graph representation model on the link prediction and knowledge base QA tasks. More research is needed to examine their performance on information retrieval, recommendation system and other tasks. [Conclusions] There are significant differences between the translation-based and the semantic matching-based knowledge graph representation models. The score function, negative sampling, and optimization method of the knowledge graph representation model, as well as the proportion of training data have significant impacts on the results of the link prediction.