CoKE: Contextualized Knowledge Graph Embedding

• DOI: 10.48550/arxiv.1911.02168
• Publication Date: 2019-01-01
• Authors: Quan Wang; Pingping Huang; Haifeng Wang; Songtai Dai; Wenbin Jiang; Jing Liu; Yajuan Lyu; Yong Zhu; Hua Wu
• Keywords: Knowledge graph

Knowledge graph embedding, which projects symbolic entities and relations into continuous vector spaces, is gaining increasing attention. Previous methods allow a single static embedding for each entity or relation, ignoring their intrinsic contextual nature, i.e., may appear in different contexts, accordingly, exhibit properties. This work presents Contextualized Graph Embedding (CoKE), novel paradigm that takes account such learns dynamic, flexible, fully contextualized relation embeddings. Two types of contexts are studied: edges paths, both formulated as sequences relations. CoKE sequence input uses Transformer encoder to obtain representations. These representations hence naturally adaptive the input, capturing meanings therein. Evaluation on wide variety public benchmarks verifies superiority link prediction path query answering. It performs consistently better than, at least equally well current state-of-the-art almost every case, particular offering an absolute improvement 21.0% H@10 Our code available \url{https://github.com/PaddlePaddle/Research/tree/master/KG/CoKE}.