Dual-Axis Compositional Contrastive Few-Shot Learning using Prototypes Across Linguistic and Semantic Dimensions for Indic Low-Resource Multilingual NLU

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Dual-Axis Compositional Few-Shot Learning (DAC-FSL) is a new framework addressing the challenge of adapting Multilingual Natural Language Understanding (NLU) systems to new languages or semantic labels with limited annotated examples, particularly for low-resource Indic languages. Unlike conventional models that entangle linguistic and semantic representations, DAC-FSL explicitly factorizes the representation space into independent linguistic and semantic embedding axes. It constructs joint representations compositionally through multiplicative interaction of these axis-specific embeddings. The framework integrates factorized prototype learning, axis-structured contrastive alignment, and disentanglement regularization using HSIC-based statistical independence and Jacobian-based cross-axis decorrelation. Experiments on six low-resource Indic languages (Hindi, Bengali, Sanskrit, Assamese, Tamil, Telugu) demonstrated strong performance, achieving 81.12% accuracy for few-shot languages with known semantics, 63.5% for new semantic classes from few-shot examples, and 89.56% on known language and seen semantics. This approach enables stable compositional generalization for scalable multilingual NLU.

Key takeaway

For NLP Engineers building multilingual NLU systems for low-resource languages, you should consider Dual-Axis Compositional Few-Shot Learning (DAC-FSL). This framework explicitly factorizes linguistic and semantic representations, offering a stable approach to adapt models when new languages or semantic labels are introduced with limited data. Explore implementing axis-factorized architectures to improve compositional generalization and scalability in linguistically diverse environments.

Key insights

Factorizing linguistic and semantic representations enables robust few-shot adaptation in multilingual NLU, especially for low-resource languages.

Principles

• Factorize linguistic and semantic embedding axes.
• Construct joint representations compositionally.
• Apply disentanglement regularization for stability.

Method

Dual-Axis Compositional Few-Shot Learning factorizes linguistic and semantic embedding axes. It uses factorized prototype learning, axis-structured contrastive alignment, and disentanglement regularization with HSIC and Jacobian-based decorrelation.

In practice

• Implement axis-factorization for low-resource NLU.
• Evaluate compositional models for evolving label spaces.
• Apply HSIC/Jacobian for representation disentanglement.

Topics

• Multilingual NLU
• Few-Shot Learning
• Low-Resource Languages
• Representation Learning
• Contrastive Learning
• Indic Languages

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Rethinking the Good Enough Embedding for Easy Few-Shot Learning — High-quality, off-the-shelf embeddings are sufficient for state-of-the-art few-shot classification, bypassing complex meta-learning.
• Multilinguality at the Edge: Developing Language Models for the Global South — Multilinguality and edge deployment intersect at the "last mile" challenge for equitable LM access in the Global South.
• Modular Monolingual Adaptation using Pretrained Language Models — Modular PLM adaptation, freezing embeddings while tuning the rest, improves low-resource language performance.
• 1S-DAug: One-Shot Data Augmentation for Robust Few-Shot Generalization — 1S-DAug improves few-shot learning generalization by generating diverse, faithful image variants at test time via diffusion.
• Supervision versus Demonstration-Based In-Context Learning for Multiword Expression Classification — In-context learning for Turkish LVCs is highly prompt-sensitive, with few-shot demonstrations improving LLM calibration to match supervised baselines.
• Multi-hop Relational Contrastive Learning: Extending Spatial Contrastive Pre-training Beyond Pairwise Relations — Multi-hop relational contrastive learning captures complex spatial dependencies beyond pairwise relations, improving scene understanding.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.