Robust Zero-Shot Intent Detection via Contrastive Transfer Learning

Intent detector is a central component of any task-oriented conversational system. The goal of the intent detector is to identify the user's goal by classifying natural language utterances. In recent years, research has focused on supervised intent detection models. Supervised learning approaches cannot accommodate unseen intents, which may emerge after the system has been deployed - the more practically relevant setting, known as zero-shot intent detection. The existing zero-shot learning approaches split a dataset into seen and unseen intents for training and evaluations without taking the sensitivity of the data collection process into account. That is, humans tend to use repeated vocabulary and compose sentences with similar compositional structures. We argue that the source-to-target relationship learning objective of zero-shot approaches under typical data split procedure renders the zero-shot models prone to misclassifications when target intents are divergent from source intents. To this end, we propose INTEND, a zero-shot INTENt Detection methodology that leverages contrastive transfer learning and employs a zero-shot learning paradigm in its true sense. First, in contrast to partitioning the training and testing sets from the same dataset, we demonstrate that selecting training and testing sets from two different datasets, allows for rigorous zero-shot intent detection evaluations. Second, our employed contrastive learning goal encourages the system to focus on learning a generic similarity function, rather than on commonly encountered patterns in the training set. We conduct extensive experimental evaluations using four public intent detection datasets for up to 150 unseen classes. Our experimental results show that INTEND consistently outperforms state-of-the-art zero-shot techniques by a substantial margin. Furthermore, our approach achieves significantly better performance than few-shot intent detection models.