Enhance the Saliency: Synthesize Text Noise Samples for Few-Shot Out-of-Distribution Intent Detection

Few-shot out-of-distribution (OOD) detection is a critical yet under explored scenario in dialogue systems. Existing data augmentation techniques either incorporate external data or generate hard negative samples within the feature space, which often leads to issues such as introducing knowledge bias, failing to align with the discrete nature of text, and inadequately addressing the problem of under-representation caused by in-distribution (IND) overfitting. Motivated by the recent findings that enhancing intra-class discrimination can mitigate IND overfitting, and the class of a sentence is predominantly determined by salient words, we propose EnSal, a method designed to strengthen the features of salient words in order to enhance the correlation between intent features and their corresponding classes. To achieve this, we jointly train k-nearest neighbors contrastive learning (KCL) alongside cross-entropy (CE) to improve the intra-class discrimination of intent features. Salient words are identified using both the k-nearest neighbors condition and the prediction probability condition. These words are retained as templates for synthesizing text samples, thereby avoiding the introduction of knowledge bias while preserving consistency with the discrete characteristics of text. Furthermore, we treat the synthetic text as noise samples associated with their corresponding training samples and perform denoising autoencoder (DAE) training on the augmented dataset. This process enables the identification of common and significant class features, effectively alleviating the under-representation issue. Extensive experimental results demonstrate that our method surpasses the current state-of-the-art in few-shot OOD intent detection.