Contrastive Learning-Enhanced Nearest Neighbor Mechanism for Multi-Label Text Classification

Multi-Label Text Classification (MLTC) is a fundamental and challenging task in natural language processing. Previous studies mainly focus on learning text representation and modeling label correlation. However, they neglect the rich knowledge from the existing similar instances when predicting labels of a specific text. To address this oversight, we propose a k nearest neighbor (kNN) mechanism which retrieves several neighbor instances and interpolates the model output with their labels. Moreover, we design a multi-label contrastive learning objective that makes the model aware of the kNN classification process and improves the quality of the retrieved neighbors during inference. Extensive experiments show that our method can bring consistent and considerable performance improvement to multiple MLTC models including the state-of-the-art pretrained and non-pretrained ones.