基于对比学习的临床领域意图识别算法研究

Contrastive Learning-Based Algorithm for Clinic Intent Recognition

随着电子信息化的发展,智能搜索、知识问答等应用被越来越多地应用在临床领域中.意图识别作为其中重要的一部分,随着这类应用的逐渐兴起,受到越来越多的关注.意图识别即理解用户问句的意图.在自然语言处理中,意图识别的本质是文本分类问题.针对意图识别工作,大量的研究和探索用以理解用户的文本输入,并将其映射到预先给定的意图类别中.本文提出一种基于对比学习的意图识别算法,根据文本的长度和意图类别的数量,将意图识别定义为短文本多分类问题.通过将对比学习引入到分类模型的训练中,提高模型的最终效果.在有监督学习场景中,采用R-drop对比学习方法.该方法选择dropout作为数据增强的方式,同时通过增加一个正则项来强化模型对dropout的鲁棒性.同时,对数据进行无监督训练,作为预训练阶段.并在预训练过程中选择SimCSE对比学习方法.最终将无监督学习与有监督学习结合,形成基于半监督学习的R-SimCSE模型.实验选取textCNN、textRNN、textRCNN、BERT-base、prompt等多种分类模型进行对比.实验结果显示,基于对比学习的分类模型效果优于文中选择的其他分类算法模型,在CHIP-QIC数据集上,准确率提升了0.0097~0.0493.

With the development of electronic informatization,applications such as intelligent search and knowledge Q&A are being increasingly explored in the clinical field.As a crucial element of such applications,intent recognition has received immense attention.Intent recognition involves the understanding of the intent of user questions.In natural language processing,intent recognition is a text classification problem wherein substantial research and exploration are conducted to understand the users’text inputs and map them into the prespecified intent categories.This paper proposes an intent recognition algorithm based on contrastive learning,which defines intent recognition as a short text multiclassification problem based on the length of the text and the number of intent categories.The effect has been enhanced using contrastive learning in the classification model training.Furthermore,in the supervised learning scenario,the R-drop contrastive learning method is adopted.This method chooses dropout for data enhancement and increases the robustness of the model to dropout by adding a regular term.Concurrently,unsupervised training is performed as a pretraining stage.The SimCSE contrastive learning method is chosen in the pretraining process.Finally,the combination of unsupervised and supervised learning forms the R-SimCSE model based on semisuper-vised learning.Moreover,several classification models are selected for comparison in the experiment,such as textCNN,textRNN,textRCNN,BERT-base,and prompt.The result shows that the classification model based on contrastive learning is superior to the other selected classification algorithm models.On the CHIP-QIC datasets,the accuracy rate is enhanced by 0.0097-0.0493.