相近尺寸下生成式模型和判别式模型在文本多分类任务上的性能比较研究

A Comparative Study on the Performance of Generative and Discriminative Models in Text Multi-Classification Tasks under Similar Model Sizes

摘要: 针对自然语言处理中应用广泛的文本多分类任务，为了探究其在实际落地过程中最合适的模型选型。本文从预测准确率和输出响应时延两个维度，对尺寸大小相近的生成式模型Qwen3-0.6b和判别式模型Bert-base在文本多分类任务上的性能表现进行了对比研究。作者使用了THUCNews新闻数据集，设置了Bert微调、Qwen3-0.6b零样本提示、Qwen3-0.6b全参微调和Qwen3-0.6bLoRA微调四组对比实验。实验表明，在相同训练样本的条件下，判别式模型在宏观平均F1值和预测速度上都优于最好的生成式模型方案，分别高出了0.8%和356.8%。

Abstract: In the widely applied text multi-classification task of natural language processing, this paper aims to explore the most suitable model selection for practical implementation. From the perspectives of prediction accuracy and response latency, we conducted a comparative study on the performance of the generative model Qwen3-0.6b and the discriminative model Bert-base in text multi-classification tasks. The experiments were based on the THUCNews news dataset, with four comparison groups set up: Bert fine-tuning, Qwen3-0.6b zero-shot prompting, Qwen3-0.6b full-parameter fine-tuning, and Qwen3-0.6b LoRA fine-tuning. The results demonstrated that, under the same training data conditions, the discriminative model outperformed the best-performing generative model approach in both macro-average F1 score and prediction speed, achieving improvements of 0.8% and 356.8%, respectively.