基于距离损失和决策边界的开放意图检测方法  

Open intent detection based on distance loss and decision boundary

作　　者：张盼盼 华宇[1] 勾智楠 池云仙 高凯[1] ZHANG Panpan;HUA Yu;GOU Zhinan;CHI Yunxian;GAO Kai(School of Information Science and Engineering,Hebei University of Science and Technology,Shijiazhuang,Hebei 050018,China;School of Management Science and Information Technology,Hebei University of Economics and Business,Shijiazhuang,Hebei 050061,China)

机构地区：[1]河北科技大学信息科学与工程学院,河北石家庄050018 [2]河北经贸大学管理科学与信息工程学院,河北石家庄050061

出　　处：《河北科技大学学报》2024年第6期618-626,共9页Journal of Hebei University of Science and Technology

基　　金：国家自然科学基金(61772075);河北省自然科学基金(F202208006,F2023207003);河北省高等学校科学研究青年基金项目(QN2024196)。

摘　　要：针对开放意图检测任务中对特征分布处理不足有可能导致所得特征分布不够紧凑的问题,提出一种融合BERT、距离损失、决策边界的开放意图检测方法。首先,通过BERT模型捕获文本间的上下文特征;然后,通过距离损失令样本特征学习更为紧密;最后,进行决策边界学习,实现开放意图检测任务。结果表明,所提方法在公开数据集StackOverflow上具有较好的表现,在2种不同的已知意图比例设置下表现均为最好,准确率达到88.28%和84.43%,F1值达到87.51%和87.40%。研究结果补充了针对边界检测的特征表示再处理方法,可为解决开放意图检测问题提供参考。In order to solve the problem that the feature distribution is not compact enough due to insufficient processing of feature distribution in open intent detection task,an open intent detection method integrating BERT,distance loss and decision boundary was proposed.Firstly,the context features between texts were captured by BERT model.Then,the learning of sample features was made more compact by distance loss.Finally,decision boundary learning was carried out to achieve the task of open intent detection.The results show that the proposed method has high performance on the public dataset StackOverflow,with the best performance under two different known intent ratio settings,achieving the accuracy of 88.28%and 84.43%,and the F1 values of 87.51%and 87.40%,respectively.The research results complement the future representation reprocessing method for boundary detection,and can provide reference for open intent detection.

关 键 词：自然语言处理 意图识别 意图检测 距离损失 决策边界 

分 类 号：TN391[电子电信—物理电子学]