Boosting Knowledge Base Automatically via Few-Shot Relation Classification

Abstract

Relation classification (RC) aims at extracting structural information, i.e., triplets of two entities with a relation, from free texts, which is pivotal for automatic knowledge base construction. In this paper, we investigate a fully automatic method to train a RC model which facilitates to boost the knowledge base. Traditional RC models cannot extract new relations unseen during training since they define RC as a multiclass classification problem. The recent development of few-shot learning (FSL) provides a feasible way to accommodate to fresh relation types with a handful of examples. However, it requires a moderately large amount of training data to learn a promising few-shot RC model, which consumes expensive human labor. This issue recalls a kind of weak supervision methods, dubbed distant supervision (DS), which can generate the training data automatically. To this end, we propose to investigate the task of few-shot relation classification under distant supervision. As DS naturally brings in mislabeled training instances, to alleviate the negative impact, we incorporate various multiple instance learning methods into the classic prototypical networks, which can achieve sentence-level noise reduction. In experiments, we evaluate our proposed model under the standard N-way K-shot setting of few-shot learning. The experiment results show that our proposal achieves better performance.