A new paper proposes an automated fix for a boring but real problem: messy training labels for named entity recognition, the task of teaching software to spot people, places, and organizations in text.
The researchers built a multi-step pipeline that repeatedly re-labels a NER dataset using self-training, a frequency-based approach that tracks how often the model's guesses agree across iterations. A dual-threshold mechanism sets separate confidence bars for accepting or rejecting a label, which the authors say sharpens the model's certainty over each pass. Tested on several NER datasets, the corrected versions produced measurably better performance than the originals. The paper also tries generative large language models as annotators for low-resource languages, where clean labeled data is scarce.
That last point is the real story. Most NER research assumes reasonably clean training data, an assumption that falls apart for languages without large annotation budgets or established corpora. A tool that automatically scrubs noisy labels could matter more for a low-resource language than for English, where good NER datasets already exist.
Worth noting: the performance gains come from the authors' own benchmarks, and self-training methods have a known failure mode - they can reinforce whatever biases already exist in the seed labels rather than removing them.