This article investigates document-level relation extraction (DocRE) in low-resource settings. The authors identified that existing models trained on scant data overestimate the ‘no relation’ label, limiting performance. To address this, they proposed PRiSM, a method that calibrates scores based on relation semantic information. The evaluation of PRiSM on three DocRE datasets showed significant performance improvement and reduction in calibration error even when trained with about 3% of data.
Publication date: 25 Sep 2023
Project Page: https://github.com/brightjade/PRiSM
Paper: https://arxiv.org/pdf/2309.13869
