Reducing Sentiment Bias in Pre-trained Sentiment Classification via Adaptive Gumbel Attack

Pre-trained language models (PLMs) have recently enabled rapid progress on sentiment classification under the pre-train and fine-tune paradigm, where fine-tuning phase aims to transfer factual knowledge learned by PLMs classification. However, current methods ignore risk that cause problem of bias, is, tend inject positive or negative from contextual information certain entities (or aspects) into their word embeddings, leading them establish spurious correlations with labels. In this paper, we propose an adaptive Gumbel-attacked classifier immunes bias adversarial-attack perspective. Due complexity diversity construct multiple Gumbel-attack expert networks generate various noises mixed Gumbel distribution constrained mutual minimization, design training framework synthesize complex noise confidence-guided controlling number networks. Finally, capture these effectively simulate based feedback classifier, then a multi-channel parameter updating algorithm strengthen recognize fusing parameters between each network. Experimental results illustrate our method significantly reduced improved performance