The article discusses a study on Unsupervised Relation Extraction (URE), a process that extracts relations between named entities from raw text without manual annotations. The study highlights the importance of contrastive learning strategies, diverse positive pairs, and appropriate loss functions in URE. The authors propose AugURE, a method that uses within-sentence pairs augmentation and cross-sentence pairs extraction to increase the diversity of positive pairs. They also suggest replacing noise-contrastive estimation (NCE) loss with margin loss for better relation representation learning. The proposed method reportedly achieved state-of-the-art performance on NYT-FB and TACRED datasets.
Publication date: 4 Dec 2023
Project Page: https://github.com/qingwang-isu/AugURE
Paper: https://arxiv.org/pdf/2312.00552
