Ensemble Dual-Channel Model for Low-Resource Text Classification

Low-resource text classification is sensitive to computational and memory resources. Model ensemble combines outputs from diverse base learners to achieve superior performance. Current research in Natural Language Processing predominantly focuses on heterogeneous ensemble (HEE) methods using BERT-related models, which integrate different architectures at the cost of substantial computational and memory resources. In this work, we propose Ensemble Dual-Channel (EDC) Model of Kolmogorov-Arnold Networks (KAN) and Multilayer Perceptron (MLP) for text classification that combines homogeneous and heterogeneous ensembles, specifically designed for low-resource scenarios. The EDC model significantly reduces computational and memory resource consumption by adopting homogeneous ensemble methods. HEE is implemented through the KAN-MLP dual-channel model, which requires negligible additional parameters and training time while significantly improving model diversity. We conduct experiments on five publicly available datasets, and the classification accuracy of the EDC model improves by at least 0.1% and up to 1.46%. The experimental results demonstrate the effectiveness of the EDC model.