基于大模型标签蒸馏的搜索意图识别OA
Search Intent Recognition Based on Large Model Label Distillation
在搜索引擎中,准确识别用户查询的意图对提升搜索体验至关重要.搜索意图识别属于短文本分类任务,传统方法依赖大量人工标注数据,成本高昂且难以适应新意图的快速涌现.文章提出的基于大模型标签蒸馏的搜索意图识别方法,利用大语言模型(如GPT4o、DeepSeek-R1、星火x1)的强大语义理解能力,为无标签查询指令生成高质量意图标签,构建训练数据集;进而通过知识蒸馏技术,将大模型的知识迁移至轻量级预训练模型(如ERNIE 3.0、BERT)进行微调.实验结果表明,该方法在13.6万规模的中文数据集上显著提升了模型性能,在降低标注成本的同时,有效提升了意图识别效率.
In search engines,accurately recognizing the intent of user queries is crucial for improving search experience.Search intent recognition belongs to the task of short text classification.Traditional methods rely on massive manually labeled data,which implies high costs and difficulty in adapting to the rapid emergence of new intents.This paper proposes a search intent recognition method based on large model label distillation.It utilizes the powerful semantic understanding capabilities of Large Language Models(such as GPT4o,DeepSeek-R1,and Spark x1)to generate high-quality intent labels for unlabeled query instructions and construct training datasets.Furthermore,through Knowledge Distillation technology,the knowledge of large models is transferred to lightweight pre-trained models(such as ERNIE 3.0 and BERT)for fine-tuning.Experimental results show that this method significantly improves model performance on a Chinese dataset with a scale of 136 000,and effectively enhances intent recognition efficiency while reducing labeling costs.
李睿琪;秦志鹏
科大讯飞股份有限公司,安徽 合肥 230088安徽皖通科技股份有限公司,安徽 合肥 230088
信息技术与安全科学
意图识别文本分类标签蒸馏大模型预训练模型
intent recognitiontext classificationlabel distillationlarge modelpre-trained model
《现代信息科技》 2026 (3)
40-44,5
评论