Meta Self-Refinement for Robust Learning with Weak Supervision

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1043–1058 May 2-6, 2023 ©2023 Association for Computational Linguistics Meta Self-Refinement for Robust Learning with Weak Supervision Dawei Zhu1, Xiaoyu Shen1, Michael A. Hedderich1,2, Dietrich Klakow1 1Saarland University, Saarland Informatics Campus, Germany 2Cornell University, United States {dzhu,xshen,mhedderich,dietrich.klakow}@lsv.uni-saarland.de Abstract Training deep neural networks (DNNs) under weak supervision has attracted increasing re- search attention as it can significantly reduce the annotation cost. However, labels from weak supervision can be noisy, and the high capacity of DNNs enables them to easily overfit the label noise, resulting in poor generalization. Recent methods leverage self-training to build noise- resistant models, in which a teacher trained un- der weak supervision is used to provide highly confident labels for teaching the students. Nev- ertheless, the teacher derived from such frame- works may have fitted a substantial amount of noise and therefore produce incorrect pseudo- labels with high confidence, leading to severe error propagation. In this work, we propose Meta Self-Refinement (MSR), a noise-resistant learning framework, to effectively combat label noise from weak supervision. Instead of rely- ing on a fixed teacher trained with noisy labels, we encourage the teacher to refine its pseudo- labels. At each training step, MSR performs a meta gradient descent on the current mini-batch to maximize the student performance on a clean validation set. Extensive experimentation on eight NLP benchmarks demonstrates that MSR is robust against label noise in all settings and outperforms state-of-the-art methods by up to 11.4% in accuracy and 9.26% in F1 score. 1 Introduction Fine-tuning pre-trained language models has led to great success across NLP tasks. Nonetheless, it still requires a substantial amount of manual labels to achieve satisfying performance on many tasks. In reality, obtaining large amounts of high-quality labels is costly and labor-intensive (Davis et al., 2013). For certain domains, it is even infeasible due to legal issues and lack of data or domain experts. Weak supervision is a widely-used approach for reudcing such cost by leveraging labels from weak sources, e,g., heuristic rules, knowledge bases or lower-quality inexpensive crowdsourcing (Ratner et al., 2017; Zhou et al., 2020; Lison et al., 2020). It has raised increasing attention in recent years, and efforts have been made to quantify the progress on weakly supervised learning, like the WRENCH benchmark (Zhang et al., 2021). Although weak labels are inexpensive to ob- tain, they are often noisy and inherit biases from weak sources. Training neural networks with weak labels is challenging because of their immense capacity, which leads them to heavily overfit to the noise distribution, resulting in inferior gener- alization performance (Zhang et al., 2017). Vari- ous approaches have been proposed to tackle this challenge. Earlier research focused primarily on simulated noise (Bekker and Goldberger, 2016; Hendrycks et al., 2018), required prior knowl- edge (Ren et al., 2020; Awasthi et al., 2020) or relied on context-free aggregation rules without leveraging modern pre-trained language models (Ratner et al., 2017; Fu et al., 2020). Recently, Yu …

Content truncated to 500 words