Single-to-Multiple: Learning Multiple Intent Detection with Only Single Intent Data

Multi-intent detection can identify multiple intents of users, which gains increasing attention. However, compared to single-intent user data, obtaining high-quality and realistic multi-intent data is challenging and costly, thereby hindering the development of multi-intent detection models. To tackle the issue of limited data availability, in this work, we introduce a novel and important task called single-to-multiple intent detection (S2MID) to learn multiple intent detection with only single-intent data, which can greatly alleviate the multi-intent data annotation burden. To achieve this, we investigate three paradigms to address this task including: (1) <italic>Generativeprediction Paradigm</italic>; (2) <italic>Classification-based Paradigm</italic>; and (3) <italic>Text-match Paradigm</italic>. Experimental results demonstrate that the introduced paradigms can generate multiple intents with only single intent data. In addition, extensive analyses reveal that S2MID task is challenging, even for powerful large-scale language models. Finally, we summarize the challenges and point out some future works, hoping to provide new insights to spur more breakthroughs in S2MID.