Semi-supervised Learning for Fine-Grained Entity Typing with Mixed Label Smoothing and Pseudo Labeling

Abstract

Distant supervision (DS) has been proposed to automatically annotate data and achieved significant success in fine-grained entity typing(FET). Despite its efficiency, distant supervision often suffers from the noisy labeling problem. To solve the noisy labeling problem, existing approaches assume the existence of “clean” and “noisy” sets in the training data and use different types of methods to utilize them. However, they still suffer from the confirmation bias problem in the “noisy” set and the false positive problem in the “clean” set. To address these issues, we propose a novel semi-supervised learning method with mixed label smoothing and pseudo labeling for distantly supervised fine-grained entity typing. Specifically, to solve the false positive problem on the “clean” set, we propose a mixed label smoothing method to smooth the labels of the “clean” set to train the FET model. To solve the confirmation bias problem on the “noisy” set, we do not consider the labels in the “noisy” set and use a pseudo labeling technique to deal with the “noisy” set. Extensive experiments conducted on three widely used FET datasets show the effectiveness of our proposed approach. The source code is publicly available at https://github.com/xubodhu/NFETC-SSL.