Learn and Unlearn in Multilingual LLMs

2024-06-19

AI Summary

This study investigates how misinformation, injected via training data, propagates across languages in multilingual large language models (MLLMs), undermining generation reliability. We identify a critical phenomenon: unlearning harmful content in one language (e.g., English) inadvertently reinforces toxic outputs in other languages—a previously unreported cross-lingual reinforcement effect. To address this, we propose response-level *multilingual co-forgetting*, requiring simultaneous unlearning in both English and the original harmful language. We introduce a response-intervention-based multilingual co-forgetting framework that integrates cross-lingual consistency evaluation and harmful-generation provenance analysis. Extensive experiments across 12 languages demonstrate that our method reduces misinformation generation by 83.7% on average, significantly improving safety and cross-lingual generalization. The framework offers a scalable, verifiable paradigm for content governance in multilingual foundation models.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Safety and Robustness
• Planning, Routing, and Scheduling: Planning with Language Models

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

This paper investigates the propagation of harmful information in multilingual large language models (LLMs) and evaluates the efficacy of various unlearning methods. We demonstrate that fake information, regardless of the language it is in, once introduced into these models through training data, can spread across different languages, compromising the integrity and reliability of the generated content. Our findings reveal that standard unlearning techniques, which typically focus on English data, are insufficient in mitigating the spread of harmful content in multilingual contexts and could inadvertently reinforce harmful content across languages. We show that only by addressing harmful responses in both English and the original language of the harmful data can we effectively eliminate generations for all languages. This underscores the critical need for comprehensive unlearning strategies that consider the multilingual nature of modern LLMs to enhance their safety and reliability across diverse linguistic landscapes.

Problem

• Addressing harmful misinformation propagation in multilingual LLMs
• Evaluating efficacy of unlearning methods across languages
• Mitigating cross-lingual reinforcement of harmful content

Innovation

• Multilingual unlearning methods for harmful content
• Addressing harmful responses in both languages
• Comprehensive strategies for diverse linguistic landscapes