Improving distantly supervised named entity recognition by emphasizing uncertain examples

Distantly supervised named entity recognition (DS-NER) aims to acquire knowledge from noisy labels. Recently, label re-weighting and label correction based frameworks have been recognized as promising approaches for DS-NER. These methods mainly handle easy or hard examples, yet neglect the impact of uncertain examples that are predicted correctly sometimes and incorrectly some other times during optimization. In this paper, we propose UE-NER, an Uncertainty Estimation method for DS-NER, which estimates the uncertainty of training examples and emphasizes uncertain ones, thus leads to more accurate and robust performance. To enable uncertainty reasoning, we formulate DS-NER as a span-level classification problem and the variance in predicted probability of the correct class across iterations of minibatch SGD is taken as the uncertainty measure. We further design an enhanced encoder to combine the power of the named entity and other spans in the sentence to boost recognition performance. Experimental results on two benchmark datasets demonstrate the superiority of the proposed UE-NER over existing DS-NER methods.