1+1>2: Can Large Language Models Serve as Cross-Lingual Knowledge Aggregators?

TL;DR

This paper proposes a novel method to improve multilingual performance of Large Language Models by aggregating cross-lingual knowledge, reducing language disparities, and enhancing consistency across languages.

Contribution

It introduces a comprehensive knowledge aggregation approach with language-specific detectors and answer integration, advancing the multilingual capabilities of LLMs.

Findings

Significant performance improvements in multilingual tasks

Reduction in language performance disparity

Each component of the method contributes to overall enhancement

Abstract

Large Language Models (LLMs) have garnered significant attention due to their remarkable ability to process information across various languages. Despite their capabilities, they exhibit inconsistencies in handling identical queries in different languages, presenting challenges for further advancement. This paper introduces a method to enhance the multilingual performance of LLMs by aggregating knowledge from diverse languages. This approach incorporates a low-resource knowledge detector specific to a language, a language selection process, and mechanisms for answer replacement and integration. Our experiments demonstrate notable performance improvements, particularly in reducing language performance disparity. An ablation study confirms that each component of our method significantly contributes to these enhancements. This research highlights the inherent potential of LLMs to harmonize multilingual capabilities and offers valuable insights for further exploration.