Robust Adaptive Feature Enhancement and Contrastive Clustering for New Intent Discovery

The primary goal of the New Intent Discovery (NID) algorithm is to discover specific intentions from user queries. State-of-the-art approaches currently focus on utilizing pre-trained language models. They depend heavily on labeled utterances and employ pseudo-labeling methods for representation learning and clustering. However, there is a lack of consideration in key tokens, which can strongly influence the semantic of other tokens and sentences. In this paper, we propose a pre-trained feature enhancement and neighbors-based contrastive clustering (PFE-NBCC) specifically for NID. This method combines a feature enhancement extractor to strengthen key features and contrastive learning to optimize the cluster of key knowledge. Firstly, we focus on identifying and processing essential features to ensure features can be recognized and utilized. Then, we apply these features to enrich the dataset in the contrastive learning method, leading to better clustering results. Experiments on three widely used benchmarks demonstrate that our method outperforms current existing methods and reaches a new state-of-the-art level in both unsupervised and semi-supervised scenarios.