Integrated embedding approach for knowledge base completion with CNN

Abstract

In knowledge graph embedding, sophisticated models may suffer from over-fitting and high computational costs. On the contrary, transitional models come with lower complexity but struggle with complex relations while integrating relational attributes or semantic information could help with embedding representation. Convolutional neural networks employed in recent researches are able to model interactions between entities and relations efficiently but may ignore global dependencies. To tackle such problems, a model called Integrated Embedding Approach for Knowledge Base Completion (IEAKBC) is proposed. In this model, embedding representations of entities and relations are put together to constitute a three-column, k dimensional matrix for each triplet. Afterwards, features from different relations are integrated into head and tail entities thus forming fused triplet matrices. Both sets of matrices are used as inputs to a convolutional neural network (CNN) framework. In CNN, kernels go over each row of the matrices for feature extraction. Feature maps are subsequently concatenated and weighted for output scores to discern whether the original triplet holds or not. Experiments on four benchmark datasets show that our model performs well on complex relations while retaining transitional characteristics. Finally, we apply the model to a personalized search application, verifying its practicality in real-world scenarios.