Modeling Discriminative Representations for Out-of-Domain Detection with Supervised Contrastive Learning

摘要

Detecting Out-of-Domain (OOD) or unknown intents from user queries is essential in a task oriented dialog system. A key challenge of OOD detection is to learn discriminative se mantic features. Traditional cross-entropy loss only focuses on whether a sample is correctly classified, and does not explicitly distinguish the margins between categories. In this pa per, we propose a supervised contrastive learn ing objective to minimize intra-class variance by pulling together in-domain intents belong ing to the same class and maximize inter-class variance by pushing apart samples from differ ent classes. Besides, we employ an adversar ial augmentation mechanism to obtain pseudo diverse views of a sample in the latent space. Experiments on two public datasets prove the effectiveness of our method capturing discrim inative representations for OOD detection.

会议
ACL 2021
曾致远
曾致远
硕士研究生

自然语言理解,文本生成

何可清
硕士研究生

对话系统,摘要,预训练

严渊蒙
严渊蒙
硕士研究生

自然语言理解,预训练

刘子君
刘子君
硕士研究生

任务导向型对话系统

吴亚楠
吴亚楠
硕士研究生

自然语言理解

徐红
硕士研究生

自然语言处理,意图识别

徐蔚然
徐蔚然
副教授,硕士生导师,博士生导师

信息检索,模式识别,机器学习