Contrastive variational representation generation for zero-shot intent classification

Abstract

Zero-shot intent classification (ZSIC) predicts intent labels for classes without labeled training utterances. In conventional zero-shot learning (ZSL), test utterances belong only to unseen intents; in generalized zero-shot learning (GZSL), held-out seen-intent utterances and unseen-intent utterances are evaluated jointly. We first establish a transparent semantic-alignment baseline, BERT-Align, and show that its very low unseen-intent GZSL score is associated with seen-label score bias rather than the absence of unseen labels from its output space. We then propose Contrastive Variational Representation Generation (CVRG), a two-phase framework that combines a conditional variational autoencoder (CVAE) with supervised contrastive learning. The CVAE regularizes class-conditioned feature variation, whereas the contrastive objective shapes class-relative geometry in the projected intent space. In the second phase, the trained CVAE decoder is conditioned on prior latent samples and seen or unseen intent embeddings to synthesize representations that refine only the projection network. Experiments on SNIPS and SMP-2018 include conventional and generalized ZSL, simple and recent PLM/prompt baselines, class-held-out semantic-fidelity diagnostics, intent-name perturbations, contrastive-strength analyses, and multiple seen–unseen class partitions. The results show that CVRG is competitive with recent prompt/generation methods and improves the seen–unseen balance on average, while also revealing two boundaries: contrastive separation may harm unseen-only transfer when semantic neighborhoods are distorted, and performance remains sensitive to low-quality intent metadata.