Research on Meta-Learning Based Few-Shot Chinese Text Classification Algorithms

Current text classification methods face limitations due to dataset size, hindering fast learning and generalization in few-shot scenarios. Our proposed method, Reptile-MAM, transforms multiclassification tasks into binary ones with limited source domain data. Using a task-agnostic MacBert model, we extract generic features and employ the lightweight fine-tuning method MAM for rapid adaptation to new tasks. Classification training, based on sentence pair implication relationships, completes parameter updates for the lightweight model. Reptile is used for initial parameter learning, enhancing generalization on new tasks by aggregating weighted updates. Validation on Chinese few-shot learning problems using public datasets demonstrates superior performance compared to previous methods, highlighting the effectiveness of our approach, especially with the integration of lightweight fine-tuning.