Few-shot relational triple extraction with nearest neighbor matching

Abstract

Current relational triple extraction approaches rely on a large amount of labeled data which is hard for many real-world applications. How to extract relational triples has become a concern in few-shot settings. At present, the researches in few-shot settings are rare on the joint extraction tasks. The main problem is that the accuracy of subject and object recognition is relatively low. To improve the accuracy of subject and object recognition, we propose an end-to-end few-shot relational triple extraction model with nearest neighbor matching. Specifically, we recognize subject and object in the sentence according to the semantic similarity of words, and propose the new method to discriminate the subject and object. Also, we add the loss of the subject and object type as a penalty item for model training. Experiments results demonstrate that our method greatly improves the performance of subject and object recognition on the public English dataset FewRel, and achieves the state-of-the-art on the few-shot relational triple extraction task.