Open Access

ARTICLE

COPA: Confidence-Guided Orthogonality-Constrained Prompt Adaptation for Few-Shot Relation Classification

Shunran Duan, Meijuan Yin^*, Xiangyang Luo, Lunchong Cui, Chenyu Wang

Henan Provincial Key Laboratory of Cyberspace Situational Awareness, Information Engineering University, Zhengzhou, China

* Corresponding Author: Meijuan Yin. Email:

Computers, Materials & Continua 2026, 89(1), 23 https://doi.org/10.32604/cmc.2026.084290

Received 20 April 2026; Accepted 16 June 2026; Issue published 13 August 2026

Abstract

Few-shot relation classification aims to identify semantic relations between entity pairs under limited annotated data. Although recent prompt learning-based methods have achieved promising performance, they often rely on manually crafted, domain-specific prompt templates, which restrict their transferability across domains. In this paper, building upon the multi-task prompt transfer paradigm of MPT, we propose a Confidence-guided Orthogonality-constraint Prompt Adaptation framework for few-shot relation classification, named COPA. The proposed framework learns a shared domain-invariant prompt matrix together with domain-specific low-rank prompt matrices via multi-domain soft prompt tuning, enabling the transfer of domain-invariant relational knowledge across domains. Unlike MPT, to explicitly disentangle domain-invariant and domain-specific information, we further introduce an orthogonality constraint that encourages the shared prompt to capture invariant relational semantics while forcing the domain-specific prompt to model complementary domain residuals. For target domain adaptation, we reuse the shared prompt as prior knowledge and combine it with a target domain-specific low-rank matrix, followed by a confidence-guided prompt adaptation strategy that encourages domain-invariant knowledge preservation while facilitating efficient adaptation to the target domain. Experiments on four public Chinese datasets demonstrate that COPA achieves consistently competitive and stable results, with 1.09%–2.08% absolute Micro-F1 improvements over the strongest MPT-based baseline in representative settings and overall gains ranging from 0.83% to 8.72% across all compared baselines.

---

Keywords

---