注释噪声在基于变换器的命名实体识别模型中的传播

In the age of large-scale language models, transformer-based architectures have significantly enhanced the performance of named entity recognition (NER) systems. Despite improvements in model capacity and contextual understanding, annotation noise remains a persistent challenge that undermines accuracy. Even minor inconsistencies in labeled datasets can cascade through transformer pipelines, resulting in systemic errors that are difficult to diagnose and correct. This article examines the origins of annotation noise, its propagation within transformer-based NER models, and strategies organizations can adopt to mitigate its impact.