Applying Curriculum Learning on Path-based Knowledge Graph Reasoning Algorithems

Abstract

In the field of knowledge graph reasoning, path reasoning based on reinforcement learning avoids using random walking methods and the inefficient search, but what follows is the false path problem. The amount of false paths is more than that of correct ones. The agent would usually reach the correct entity from the wrong paths first, and be more inclined to them in subsequent exploration. We propose to use curriculum learning to solve this problem: assuming that in the environment corresponding to the simple samples, the proportion of correct paths and the quality of paths are higher. The agent counters the sensitivity of RL models to false paths in the strategy by learning the basics knowledge out of simple sample sets. After a comprehensive evaluation on three KG datasets, our method is highly versatile and improves performance in knowledge-based question answering with almost no additional training time. And taking MINERVA as the baseline, the MRR index has increased by 1.3%, 3.7% on datasets WN18RR, NELL-995 respectively.