Ordinal PCL+: A Prompt-based Multilingual Framework for Few-shot and Zero-shot Cross-lingual Text Classification

Abstract

This paper proposes Ordinal Prompt based Cross-lingual Learning plus (Ordinal PCL+), a prompt-based multilingual framework for few-shot and zero-shot text classification. The approach inserts learnable continuous prompts, conditions token representations with a lightweight label-attention block, and performs a two-pass self-learning routine that couples pseudo-labels with representation–prototype agreement. For tasks with ordered labels, an ordinal head enforces monotonic consistency; a contrastive objective further aligns same-class instances across languages. The overall loss combines classification, agreement, ordinal, and contrastive terms and is followed by post-hoc temperature scaling to ensure well-calibrated probabilities. Evaluation on the Multilingual Amazon Reviews Corpus (MARC) uses English k-shot supervision (k {4, 8, 16, 32, 64}) with zero-shot transfer to five target languages, reporting accuracy and expected calibration error (ECE) across multiple seeds. Relative to fine-tuning and prompt-only variants, the framework consistently improves transfer while reducing calibration error and remains modular and lightweight for practical deployment on commodity hardware. The components are ablated to quantify their contributions, showing complementary gains from label attention and self-learning, consistent benefits of the ordinal branch on rating-like labels, and improved reliability after temperature scaling.