Retrieval-Based Unsupervised Noisy Label Detection on Text Data
Peiyang Liu, J. C. S. Yang, Lin Wang, Sen Wang, Yikuang Hao, Hongzhou Bai · 2023
The success of deep neural networks hinges on both high-quality annotations and copious amounts of data; however, in practice, a compromise between dataset size and quality frequently arises. Data collection and cleansing are often resource-intensive and time-consuming, leading to real-world datasets containing label noise that can introduce incorrect correlation patterns, adversely affecting model generalization capabilities. The efficient identification of corrupted patterns is indispensable, with prevalent methods predominantly concentrating on devising robust training techniques to preclude models from internalizing these patterns. Nevertheless, these supervised approaches often necessitate tailored training procedures, potentially resulting in overfitting corrupted patterns and a decline in detection performance. This paper presents a retrieval-based unsupervised solution for the detection of noisy labels, surpassing the performance of three current competitive methods in this domain.