New Intent Discovery Model Based on Data Enhancement and Contrastive Learning

• 1、School of Computer Science (National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：Task-oriented dialogue systems are widely employed in fields such as customer service support and virtual assistants. However, the majority of current task-oriented dialogue systems primarily focus on completing tasks within predefined domain intents, with less emphasis on uncovering potential new intents from user requests. This limitation constrains the intelligent evolution of dialogue systems. Therefore, this paper proposes the new intent discovery model based on data augmentation and contrastive learning (DACL) to address this issue. The model incorporates two data augmentation methods, designs corresponding loss functions, employs contrastive learning for joint training, introduces effective supervisory signals, and mitigates the isotropy of the BERT model. Through experimental validation, compared to the baseline model, this approach significantly enhances clustering performance.

Keywords： artificial intelligence dialogue system intent clustering data augmentation contrastive learning