MLRNet: A Meta-Loss Reweighting Network for Biased Data on Text Classification

Abstract

Artificially generated datasets often exhibit biases, leading conventional deep neural networks to overfit. Typically, a weighted function adjusts sample impact during model updates using weighted loss. Meta-neural networks, trained with meta-learning principles, generalize well across tasks, acquiring generalized weights. This enables the self-generation of tailored weighted functions for data biases. However, datasets may simultaneously exhibit imbalanced classes and corrupted labels, posing a challenge for current meta-models. To address this, this paper presents Meta-Loss Reweighting Network (MLRNet) with fusion attention features. MLRNet continually evolves sample loss values, integrating them with sample features from self-attention layers in a semantic space. This enhances discriminative power for biased samples. By employing minimal unbiased meta-data for guidance, mutual optimization between the classifier and the meta-model is conducted, endowing biased samples with more reasonable weights. Experiments on English and Chinese benchmark datasets including artificial and real-world biased data show MLRNet’s superior performance under biased data conditions.