GCCLA: Graph-Conditioned Cross-Lingual Adaptation of Large Language Models Under Extreme Data Scarcity (A Case Study in Tigrigna)

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

Summary

The GCCLA framework addresses the challenge of adapting large language models (LLMs) to extremely low-resource languages like Tigrigna, which suffer from severe data scarcity and a lack of structured linguistic supervision. This graph-conditioned cross-lingual adaptation method integrates multilingual knowledge graphs into parameter-efficient LLM adaptation. GCCLA conditions a frozen multilingual LLM using structured semantic and typological relations from a multilingual graph, providing a strong inductive bias for data-efficient transfer. A case study on English-to-Amharic-to-Tigrigna transfer, with 0–1000 labeled Tigrinya examples, demonstrated its effectiveness. Evaluated on five tasks—sentiment analysis, named entity recognition, natural language inference, question answering, and extractive summarization—GCCLA consistently outperformed multilingual, translation-based, and parameter-efficient baselines, achieving competitive performance with as few as 100 labeled examples and degrading gracefully under partial graph coverage.

Key takeaway

For NLP Engineers developing solutions for extremely low-resource languages, GCCLA offers a robust approach to overcome data scarcity. You should consider integrating multilingual knowledge graphs into your LLM adaptation workflows, especially when labeled data is below 1000 examples. This method significantly improves sample efficiency and performance on tasks like NER and QA, allowing you to achieve competitive results with as few as 100 labeled examples, advancing equitable NLP.

Key insights

Graph-conditioned cross-lingual adaptation using multilingual knowledge graphs enables data-efficient LLM transfer to extremely low-resource languages.

Principles

• Graph conditioning provides strong inductive bias for transfer.
• Separating knowledge representation from language modeling stabilizes learning.
• Multilingual knowledge graphs improve sample efficiency in few-shot regimes.

Method

GCCLA conditions a frozen multilingual LLM on structured semantic and typological relations from a multilingual knowledge graph, enabling data-efficient transfer to low-resource languages.

In practice

• Adapt LLMs to languages with fewer than 100 labeled examples.
• Improve few-shot learning for low-resource NLP tasks.
• Utilize existing multilingual knowledge graphs for cross-lingual transfer.

Topics

• Low-Resource Languages
• Large Language Models
• Knowledge Graphs
• Cross-Lingual Adaptation
• Parameter-Efficient Learning
• Few-Shot Learning
• Tigrigna NLP

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

Related on AIssential

• Modular Monolingual Adaptation using Pretrained Language Models — Modular PLM adaptation, freezing embeddings while tuning the rest, improves low-resource language performance.
• Cross-lingual Relation Extraction with Large Language Models: Zero-Shot, Few-Shot, and Fine-Tuned Evaluation on Romanian — QLoRA fine-tuning significantly boosts cross-lingual relation extraction performance for low-resource languages, outperforming prompt-only LLMs.
• LANG: Reinforcement Learning for Multilingual Reasoning with Language-Adaptive Hint Guidance — LANG enhances multilingual LLM reasoning by using language-adaptive hints with a progressive decay schedule, preventing language drift.
• Towards Truly Multilingual ASR: Generalizing Code-Switching ASR to Unseen Language Pairs — Code-switching ASR generalization to unseen language pairs remains limited despite model merging and domain generalization efforts.
• Specializing a Small Language Model for Closed-Domain Portuguese RAG using Knowledge Graph Supervision — Fine-tuning SLMs with knowledge graph supervision significantly enhances closed-domain RAG performance, surpassing larger LLMs.
• LEAF: Growing Trees Without Branching for Speech-Aware Large Language Model Post-Training — LEAF is a retrospective tree-based RL method for speech-aware LLM post-training that refines credit assignment.

Open in AIssential →

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