Liberating Seen Classes: Boosting Few-Shot and Zero-Shot Text Classification via Anchor Generation and Classification Reframing
Han Liu, Siyang Zhao, Xiaotong Zhang, Feng Zhang, Wei Wang, Fenglong Ma, Hongyang Chen, Hong Yu, Xianchao Zhang
TL;DR
The paper tackles the challenge of few-shot and zero-shot text classification by eliminating reliance on seen classes. It introduces a two-stage framework: (1) anchor generation, where category descriptions guide a large language model to produce diverse pseudo-samples and select representative anchors via embedding prototypes, and (2) classification reframing, which converts multi-class prediction into a binary similarity problem trained with a BERT-based classifier on anchor pairs. This approach leverages description-driven supervision and limited labeled signals to perform robust classification across diverse datasets, often outperforming strong baselines and rivaling some LLM-based methods, especially in zero-shot and few-shot regimes. The work demonstrates the practical value of anchors and binary similarity in data-scarce settings and suggests paths for extending to multi-label scenarios and broader NLP tasks.
Abstract
Few-shot and zero-shot text classification aim to recognize samples from novel classes with limited labeled samples or no labeled samples at all. While prevailing methods have shown promising performance via transferring knowledge from seen classes to unseen classes, they are still limited by (1) Inherent dissimilarities among classes make the transformation of features learned from seen classes to unseen classes both difficult and inefficient. (2) Rare labeled novel samples usually cannot provide enough supervision signals to enable the model to adjust from the source distribution to the target distribution, especially for complicated scenarios. To alleviate the above issues, we propose a simple and effective strategy for few-shot and zero-shot text classification. We aim to liberate the model from the confines of seen classes, thereby enabling it to predict unseen categories without the necessity of training on seen classes. Specifically, for mining more related unseen category knowledge, we utilize a large pre-trained language model to generate pseudo novel samples, and select the most representative ones as category anchors. After that, we convert the multi-class classification task into a binary classification task and use the similarities of query-anchor pairs for prediction to fully leverage the limited supervision signals. Extensive experiments on six widely used public datasets show that our proposed method can outperform other strong baselines significantly in few-shot and zero-shot tasks, even without using any seen class samples.