Relational learning with hierarchical attention encoder and recoding validator for few-shot knowledge graph completion

Abstract

Few-shot knowledge graph complementation (FKGC) has gained broad interest, where each task aims to complete missing facts of the long tail relations by few-shot support instances in the knowledge graph (KG). Most previous FKGC models encode few-shot relations by aggregating the neighbor information of each entity but do not make fine-grained distinctions between neighbor relations and entities. Moreover, many of those models neglect the valuable information interaction between relations, which may reduce the representational performance of the model. In this paper, we propose HARV, a novel FKGC framework based on hierarchical attention encoder and relation recoding validator. Specifically, through the hierarchical attention encoder, we can aggregate fine-grained information at the relation-level and entity-level in the neighbors respectively, thus obtaining more accurate representations. Besides, through the relation recoding validator, we are able to capture the interaction between different relations, which can decrease the over-dependence on specific entities in the few-shot relation encoding stage. Experimental results on two datasets, NELL and Wiki, demonstrate that our model performs better than the state-of-the-art baselines in different few-shot sizes. We also demonstrate the interpretability of our approach through a case study of different levels of attention for heterogeneous neighbors when encoding the central entity in the experiment.