基于对比学习的儿科问诊对话细粒度意图识别  

Fine-grained Intent Recognition from Pediatric Medical Dialogues with Contrastive Learning

作　　者：李文博 董青[2] 刘超[2] 张奇[1] LI Wenbo;DONG Qing;LIU Chao;ZHANG Qi(School of Computer Science,Fudan University,Shanghai 200433,China;The Affiliated Taian City Central Hospital of Qingdao University,Taian Shandong 271000,China)

机构地区：[1]复旦大学计算机科学技术学院,上海200433 [2]青岛大学附属泰安市中心医院,山东泰安271000

出　　处：《广西师范大学学报（自然科学版）》2024年第4期1-10,共10页Journal of Guangxi Normal University:Natural Science Edition

基　　金：国家自然科学基金面上项目(62076069)。

摘　　要：问诊对话系统的基础是自然语言理解。自然语言理解是指从对话信息中提取出意图信息和实体信息,并将其转换为结构化表达,主要包括意图识别和槽填充2种任务。意图识别是一种典型的文本分类任务,槽填充则是使用序列算法从对话文本中根据预先设定好的槽位抽取对应的槽位值。传统的方法通常对意图识别和槽填充2个任务分别构建模型,并在意图识别的基础上根据意图进行槽填充,但是这种方式容易造成错误传播。针对该问题,本文提出一种基于对比学习方法的融合对话意图分类和语义槽取值的细粒度意图识别方法。该方法结合意图分类和语义槽取值任务,使用BART作为骨干模型进行改进和创新,该模型使用编解码架构,意图识别和槽填充任务共享一个编码层,解码层采用字级别标签,通过将意图信息融合进槽填充任务,并在样本构造过程中引入对比学习。实验结果表明,本文算法在医患对话数据集上的意图识别准确率达到81.96%,槽填充的F_1分数达到85.26%,与其他算法相比有明显的效果提升。另外,通过消融实验和样例分析,进一步证明了本文算法的效果。The foundation of the inquiry dialogue system is rooted in natural language understanding(NLU),where NLU involves the extraction of intent and entity information from conversational data,transforming it into a structured representation.This process primarily encompasses two tasks:intent recognition and slot filling.Intent recognition,a typical text classification task,aims to discern the underlying purpose of the dialogue,while slot filling utilizes sequential algorithms to extract corresponding slot values based on predefined positions within the conversation.Conventional approaches often build separate models for intent recognition and slot filling,subsequently performing slot filling based on the recognized intent.However,this methodology is susceptible to error propagation.To address this issue,this paper proposes a fine-grained intent recognition method that integrates dialogue intent classification and semantic slot value extraction using a contrastive learning approach.The method combines intent classification and slot value tasks,leveraging BART as the backbone model for improvement and innovation.This model,employing an encoder-decoder architecture,shares an encoding layer for intent recognition and slot filling tasks.Additionally,it adopts character-level labels in the decoding layer,thereby integrating intent information into the slot filling task.Contrastive learning is introduced during the sample construction process.Experimental results demonstrate that the proposed algorithm achieves an intent recognition accuracy of 81.96%and a slot filling F 1 score of 85.26%on a medical dialogue dataset,showing significant performance improvements compared with other algorithms.The paper also conducts ablation experiments on contrastive learning,historical information,and sentence-level intent to further substant the effectiveness of the proposed method.

关 键 词：对比学习 意图识别 槽填充 细粒度 医疗对话 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]