Scholarlink

<p>Out-of-domain (OOD) intent detection is one of the research hotspots in task-based AI dialog. Aiming at the problems of entangling features and insufficient extraction of discriminative information, multiple feature-anisotropy regularization for OOD intent detection bidirectional encoder representations from transformer (MFAR-BERT) is proposed. The instance-level and class-level features of the intent are decoupled by a multi-head feature anisotropy disentangled strategy. To extract feature information of instance-level intent samples from different perspectives, multi-view KNN contrastive learning is designed. Correlation-matrix regularization is devised to adjust the direction of class-level sentence embeddings. Feature distributions are avoided from being confined in cone space. The experimental results show that the MFAR-BERT model achieves a minimum improvement of 0.67, 0.44, 0.49, and 1.85% on ACC_ALL, F1_ALL, F1_OOD, and F1_IND compared to DCL, SCL, ABD, HybridCL, and KNNBERT models on the publicly available datasets BANKING, OverStackflow, CLINC-SMALL, and CLINCX-FULL</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multiple feature-anisotropic regularization for out-of-domain intent detection

  • 期刊论文

  • 发表日期：

    2025年5月26日

查看全文
  • Di Wu,
  • Xiaoyu Wang,
  • Liming Feng

摘要

Out-of-domain (OOD) intent detection is one of the research hotspots in task-based AI dialog. Aiming at the problems of entangling features and insufficient extraction of discriminative information, multiple feature-anisotropy regularization for OOD intent detection bidirectional encoder representations from transformer (MFAR-BERT) is proposed. The instance-level and class-level features of the intent are decoupled by a multi-head feature anisotropy disentangled strategy. To extract feature information of instance-level intent samples from different perspectives, multi-view KNN contrastive learning is designed. Correlation-matrix regularization is devised to adjust the direction of class-level sentence embeddings. Feature distributions are avoided from being confined in cone space. The experimental results show that the MFAR-BERT model achieves a minimum improvement of 0.67, 0.44, 0.49, and 1.85% on ACC_ALL, F1_ALL, F1_OOD, and F1_IND compared to DCL, SCL, ABD, HybridCL, and KNNBERT models on the publicly available datasets BANKING, OverStackflow, CLINC-SMALL, and CLINCX-FULL