Modeling relation paths for knowledge base completion via joint adversarial training

TL;DR

This paper introduces a joint adversarial training approach using Hierarchical Attention Networks to better model relation paths for knowledge base completion, effectively combining semantic and path-based features.

Contribution

It proposes a novel method that jointly models relation paths and single relations with adversarial training, improving relation completion accuracy.

Findings

Outperforms existing path-based methods in large-scale KBs

Effectively captures shared features between relations and multi-hop paths

Model interpretability facilitates application to various relation learning tasks

Abstract

Knowledge Base Completion (KBC), which aims at determining the missing relations between entity pairs, has received increasing attention in recent years. Most existing KBC methods focus on either embedding the Knowledge Base (KB) into a specific semantic space or leveraging the joint probability of Random Walks (RWs) on multi-hop paths. Only a few unified models take both semantic and path-related features into consideration with adequacy. In this paper, we propose a novel method to explore the intrinsic relationship between the single relation (i.e. 1-hop path) and multi-hop paths between paired entities. We use Hierarchical Attention Networks (HANs) to select important relations in multi-hop paths and encode them into low-dimensional vectors. By treating relations and multi-hop paths as two different input sources, we use a feature extractor, which is shared by two downstream components (i.e. relation classifier and source discriminator), to capture shared/similar information between them. By joint adversarial training, we encourage our model to extract features from the multi-hop paths which are representative for relation completion. We apply the trained model (except for the source discriminator) to several large-scale KBs for relation completion. Experimental results show that our method outperforms existing path information-based approaches. Since each sub-module of our model can be well interpreted, our model can be applied to a large number of relation learning tasks.