A Triple-Branch Neural Network for Knowledge Graph Embedding

The distributed representation of knowledge graphs (KGs), which embeds the structured graphs into low-dimensional embedding spaces, is widely used to facilitate various applications of AI, such as information retrieval and question answering. The primary elements of KGs, the entities viewed as nodes and the relations regarded as links between entities, naturally make up the local embedding context for each other, which is called the multi-restriction property of KGs. However, this property is not fully explored by previous models, where either only part of the multi-restriction is captured, or the capability of the embedding model is limited to a fixed function without clear interpretation. To address this issue, we propose TBNN, a triple-branch neural network to learn the embeddings of KGs. In particular, the embedding of any element of a KG is determined by its multi-restriction via an interaction layer followed by parallel branched layers. Thus, the entities and relations can be treated equivalently in spite of their seeming differences in the original KG. We define the loss function of TBNN based on the confidence score of the three elements of each triple. In addition, we propose using the log-sum-exp pairwise loss to smooth the hinge loss, which results in better performance. Empirically, we evaluate our model on the tasks of link prediction and triple classification with the subsets of WordNet and Freebase. Experiment results show that our model performs better than the baselines, especially providing stable performance for relations with different mapping properties.