Language Fusion for Parameter-Efficient Cross-lingual Transfer

2025-01-12

AI Summary

Multilingual large language models (LLMs) suffer from weak cross-lingual representations and poor transfer performance due to insufficient non-English pretraining data, while existing cross-lingual adaptation methods incur substantial computational overhead. To address this, we propose a lightweight LoRA-based intra-adaptor language representation fusion method. Our approach innovatively incorporates source–target language representation alignment and interaction mechanisms directly into the low-rank linear transformation—without token mixing or parameter expansion—thereby achieving both high parameter efficiency and improved cross-lingual representation quality. Empirical evaluation on multilingual natural language understanding (NLU) tasks demonstrates significant gains: Llama 3.1 and Gemma 2 achieve +4.9% and +2.2% improvements in Exact Match on question answering, respectively, outperforming standard LoRA fine-tuning. This work introduces the first LoRA adapter architecture that enables *intra-adaptor* cross-lingual feature fusion, establishing a novel paradigm for efficient multilingual adaptation.

Technology Category

Machine Learning: Large Multimodal Models (LMMs)Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLPComputer Vision: Language and Vision

Application Category

Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web searchUser Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendationSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Limited availability of multilingual text corpora for training language models often leads to poor performance on downstream tasks due to undertrained representation spaces for languages other than English. This 'under-representation' has motivated recent cross-lingual transfer methods to leverage the English representation space by e.g. mixing English and 'non-English' tokens at the input level or extending model parameters to accommodate new languages. However, these approaches often come at the cost of increased computational complexity. We propose Fusion forLanguage Representations (FLARE) in adapters, a novel method that enhances representation quality and downstream performance for languages other than English while maintaining parameter efficiency. FLARE integrates source and target language representations within low-rank (LoRA) adapters using lightweight linear transformations, maintaining parameter efficiency while improving transfer performance. A series of experiments across representative cross-lingual natural language understanding tasks, including natural language inference, question-answering and sentiment analysis, demonstrate FLARE's effectiveness. FLARE achieves performance improvements of 4.9% for Llama 3.1 and 2.2% for Gemma~2 compared to standard LoRA fine-tuning on question-answering tasks, as measured by the exact match metric.

Problem

• Multilingual Environment
• Insufficient Training Data
• Computational Resource Consumption

Innovation

• FLARE
• Multilingual Performance Enhancement
• Cross-lingual Tasks Optimization