基于图神经网络和知识图谱的可解释小样本文本分类模型

An Interpretable Few-Shot Text Classification Model Based on Graph Neural Networks and Knowledge Graphs

摘要: 小样本文本分类具有广泛的应用场景。然而，现有方法面临两个关键挑战：数据稀缺和可解释性不足。为此，本实验提出ARExplainer方法，这是一个数据与推理增强的可解释小样本文本分类方法。通过利用大语言模型(LLMs)的泛化能力，有效扩展了训练样本的多样性，从而缓解了小样本学习的数据瓶颈。针对模型可解释性问题，构建了知识图谱驱动的推理引擎。该引擎结合图注意力网络提取可验证的符号推理路径，为分类决策提供逻辑依据，最后，利用基于提示的解释生成器生成简洁、清晰的自然语言解释。实验结果表明，在1-shot设置下，ARExplainer显著优于最好的基线模型。此外，通过与自动生成解释和人工标注结果的对比分析，证实ARExplainer能够提供更便于人类理解的自然语言解释。

Abstract: Few-shot text classification has broad application scenarios. However, existing methods face two key challenges: data scarcity and insufficient interpretability. To address these issues, this paper proposed ARExplainer, an interpretable few-shot text classification method with data and reasoning augmentation. By leveraging the generalization capability of Large Language Models (LLMs), it effectively expanded the diversity of training samples, thereby alleviating the data bottleneck in few-shot learning. For the model interpretability issue, the method constructed a knowledge graph-driven reasoning engine. This engine combined a Graph Attention Network to extract verifiable symbolic reasoning paths, providing logical evidence for classification decisions. Finally, it utilized a prompt-based explanation generator to produce concise and clear natural language explanations. Experimental results demonstrated that ARExplainer significantly outperformed the strongest baseline model in the 1-shot setting. Furthermore, comparative analysis against automatically generated explanations and human-annotated results confirmed that ARExplainer provides natural language explanations that are easier for humans to understand.