Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Yue Huang, Chenrui Fan, Yuan Li, Siyuan Wu, Tianyi Zhou, Xiangliang Zhang, Lichao Sun

TL;DR

A method to enhance the multilingual performance of LLMs by aggregating knowledge from diverse languages by incorporating a low-resource knowledge detector specific to a language, a strategic language selection process, and mechanisms for answer replacement and integration are introduced.

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.

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

TL;DR

A method to enhance the multilingual performance of LLMs by aggregating knowledge from diverse languages by incorporating a low-resource knowledge detector specific to a language, a strategic language selection process, and mechanisms for answer replacement and integration are introduced.

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.
Paper Structure (24 sections, 4 equations, 16 figures, 9 tables, 1 algorithm)

This paper contains 24 sections, 4 equations, 16 figures, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multilingual LLMs
  4. Factuality in LLMs
  5. Hallucination Mitigation
  6. Methodology
  7. Motivation
  8. Construction of Low-Resource Dataset
  9. Low-Resource Knowledge Detector
  10. Target Language Selection
  11. Answer Replacement & Integration
  12. Experiments
  13. Experiment Setup
  14. Main Results
  15. Ablation Study
  16. ...and 9 more sections

Figures (16)

  • Figure 1: The top is an example of distinct answers to the same questions in different languages. The bottom is the GPT-4's performance on 300 queries in HalluEval li2023halueval of nine different languages.
  • Figure 2: The knowledge domain of a multilingual LLM can be separated into multiple sections (the figure shows two). The language-specific knowledge (pure blue or pure orange) in one language can be utilized for improving the performance in other languages.
  • Figure 3: The average performance of six LLMs in five datasets. We show the accuracy of Chinese and English domain knowledge with the query/answer in Chinese and English.
  • Figure 4: The proposed method begins with the query detection of low-resource knowledge powered by a detector. If low-resource knowledge is detected within the queries, LLMs then select the language most likely to yield the best answer. Answer replacement and integration are employed to formulate the final response.
  • Figure 5: Statistics of the dataset in our experiments.
  • ...and 11 more figures