Divide and Denoise: Learning from Noisy Labels in Fine-grained Entity Typing with Cluster-wise Loss Correction

Fine-grained Entity Typing(FET) has witnessed great progress since distant supervision was introduced, but still suffers from label noise. Existing noise control methods applied to FET rely on predicted distribution and deals instances isolately, thus suffers from confirmation bias. In this work, We propose to tackle the two limitations with a cluster based loss correction framework named Feature Cluster Loss Correction(FCLC). FCLC first train a coarse backbone model as feature extractor and noise estimator. Then perform loss correction on each cluster to learn directly from noisy labels. Experimental results on three public datasets show FCLC achieves the best performance over existing competitive systems. Auxiliary experiments further show FCLC is stable to hyper-paramerters and even works with extream scneriaos like no clean data is available.