Tuning-Free Discriminative Nearest Neighbor Few-Shot Intent Detection via Consecutive Knowledge Transfer

Few-shot intent classification and out-of-scope (OOS) detection are core components of task-oriented dialogue systems. Solving both tasks can be challenging because of limited data availability. In this study, we aim to develop a few-shot intent classification model capable of OOS detection that does not require fine-tuning on target data. We adopt the discriminative nearest neighbor classification architecture and replace the fine-tuning phase with a consecutive pre-training approach involving natural language inference and paraphrasing tasks. Our approach leverages the training set for predictions, offering a quick and convenient way to adjust the model’s behavior by modifying a set of few labeled examples. When compared to methods that do not require fine-tuning, the developed model exhibits higher scores on various few-shot intent classification datasets.