PromptRGD: Prompt Learning with Relation-aware Gradient Denoising for Low-resource Relation Extraction

Zihang Xu, Luoyi Fu, Xinbing Wang

Abstract

Relation extraction is a fundamental task to construct a knowledge graph, which aims to find the relation categories between two entities from a string of text input sequences. Considering the fact that high-quality labeled data is rare, a recent trend is low-resource setting. Existing works either utilize prompt-based learning method to achieve encouraging results for few-shot tasks by determining an appropriate prompt template, or leverage unlabeled data by generating pseudo labels. However, the manual design of the prompt template requires not only labor-intensive but also expert knowledge. Roughly utilizing a semi-supervised method to assign pseudo-label to unlabeled data will suffer from noise accumulation. To tackle these problems, we propose a novel semi-supervised prompt learning framework with relation-aware gradient denoising for low-resource relation extraction(PromptRGD). Firstly, using learnable template words and virtual labels for prompt learning, we introduce entity type and relation prior knowledge into prompt template construction. Secondly, given the relation-aware gradient similarity between labeled and unlabeled data, PromptRGD generates a pseudo label and then improves the quality of pseudo labels on unlabeled samples in a self-training fashion. The main experimental results and a series of analyses prove the effectiveness of PromptRGD.
BibTeX
@inproceedings{xu2022promptrgd,
  author = {Zihang Xu and Luoyi Fu and Xinbing Wang},
  title = {{PromptRGD: Prompt Learning with Relation-aware Gradient Denoising for Low-resource Relation Extraction}},
  booktitle = {IEEE Big Data},
  year = {2022},
  doi = {10.1109/bigdata55660.2022.10020399},
}

Loading related content...

Figure