表示增强与双监督特征选择的畜牧业多意图识别方法OA
Augmentation and Dual-Supervised Feature Selection Method for Multi-Intent Recognition in Livestock Farming
畜牧业疾病问答是智慧畜牧服务中的重要应用场景,准确识别用户问句意图是实现疾病诊断、治疗推荐与防控知识检索的关键前提.针对畜牧业疾病问句口语化、短文本、语义稀疏和多意图共现等问题,将问句意图识别建模为多标签多意图识别任务,构建畜牧业疾病多意图问句数据集,并提出表示增强与双监督特征选择的多意图识别方法ADFS.该方法首先利用BERT获取问句上下文语义表示;其次在隐藏状态层面进行动态跨度截断与混合表示增强,提升模型对局部信息缺失和不完整表达的鲁棒性;进一步从句级嵌入空间和标签分布空间开展双监督对比学习,增强样本语义判别能力和多意图标签共现关系建模能力;最后通过特征选择模块筛选低冗余、高判别力的关键语义特征,并输入多层感知机完成多标签意图预测.实验结果表明,ADFS在畜牧业疾病多意图识别任务上取得85.26%的Macro-F1,优于多种对比模型,能够为畜牧业智能问答提供有效的问句理解支撑.
Livestock disease question answering is an important application scenario in intelligent livestock ser-vices,where accurate intent recognition is essential for disease diagnosis,treatment recommendation,and prevention knowledge retrieval.To address the colloquial expressions,short-text semantics,sparse information,and co-occurring intents in livestock disease questions,this paper formulates intent recognition as a multi-label multi-intent classifica-tion task,constructs a multi-intent dataset for livestock disease questions,and proposes an Augmentation and Dual-Supervised Feature Selection(ADFS)method for multi-intent recognition.The method first uses BERT to obtain con-textual semantic representations.Secondly,dynamic span truncation and mixed representation augmentation are per-formed at the hidden-state level to improve model's robustness against missing local information and incomplete ex-pressions.Furthermore,dual-supervised contrastive learning is further performed in both the sentence embedding space and the label distribution space to enhance semantic discrimination ability and model intent co-occurrence relationships among multi-intent labels.Finally,a feature selection module filters low-redundancy and discriminative semantic fea-tures,which are fed into a multi-layer perceptron for multi-label intent prediction.Experimental results show that ADFS achieves a Macro-F1 of 85.26%on the multi-intent recognition task for animal husbandry diseases,outperform-ing several comparison models and providing effective question understanding support for intelligent livestock question answering.
苏学谦;李世琪;张浩;刘晓宇;谢凯;刁宇峰
内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043通辽职业学院 信息技术系,内蒙古 通辽 028000内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043内蒙古民族大学 计算机科学与技术学院,内蒙古 通辽 028043
信息技术与安全科学
智慧畜牧多意图识别双监督对比学习特征选择
intelligent livestock farmingmulti-intent recognitiondual-supervised contrastive learningfeature selection
《内蒙古民族大学学报(自然科学版)》 2026 (5)
42-51,10
国家自然科学基金地区项目(62366040)国家自然科学基金青年科学基金项目(C类,62006130)中央引导地方科技发展资金项目(2026ZY0104)内蒙古自治区高等学校"青年科技英才支持计划"项目(NJYT24037)内蒙古自治区"英才兴蒙"工程个人项目暨青年拔尖人才项目(KYCYYC26002)内蒙古自治区高校青年教师科研创新能力支持项目(JYTQNXM2025-23)内蒙古自治区硕士研究生科研创新项目(NMDSSZ2631,NMDSSY2632)
评论