Multi-Agent Translation Team (MATT) - Published

Low-resource language translation project using LLM review, editing, and evaluation agents.

Abstract

This paper introduces Multi-Agent Translation Team (MATT), a workflow for refining English-to-low-resource language translations. The project tested LLM-based proofreading, editing, and evaluation roles across Vietnamese, Hindi, and Malayalam.

The workflow improved on the baseline model in the tested languages and was compared with Google Translate using automated scores and human review.

Introduction

Many online resources are still hard to use for people who do not read English. Machine translation tools such as Google Translate (GT) help, but low-resource languages remain harder because there is less training data and more room for context to be lost.

Why This Matters

  • Only 15% of the world speaks English, leaving a large share of online knowledge harder to access.
  • Countries like Vietnam and India, with growing foreign investment, still face communication barriers tied to language access.
  • In medical, legal, and research settings, poor translation can create real risk.

MATT uses separate agent roles so translation, review, editing, and scoring are handled as distinct steps.

Architecture

MATT refines translation in two stages:

  1. Initial translation generation: Uses Llama 3.1 and the Google Translate API.
  2. Iterative refinement: Uses proofreading, editing, and loss assessment agents.

Components:

  • Evaluation Coordinator: Selects the strongest base translation.
  • Proofreader and Editor: Review fluency, accuracy, style, and terminology.
  • Editor-in-Chief: Scores Loss in Translation (LiT) and handles final review.

Workflow Diagram

Results

Model Vietnamese Hindi Malayalam
Baseline 0.2471 0.1390 0.0158
MATT 0.2952 0.1681 0.0231
Google Translate 0.3886 0.2548 0.0728
  • MATT outperformed the baseline across all three tested languages.
  • Human evaluations favored MATT over both GT and baseline models for fluency and context.

Results Graph

Takeaways

  • MATT improved translation quality over the baseline in the tested low-resource settings.
  • GT remained strong on direct word-to-word scores.
  • Human reviewers preferred MATT in cases where fluency and context mattered more than literal overlap.

Limitations

  • API calls and repeated review steps add processing time.
  • LLM outputs can vary between runs, so prompts and evaluation rules need to be stable.
  • Some languages still need more training data for larger gains.

Conclusion

MATT shows one way to use agent roles for translation review: generate a draft, critique it, revise it, and score what changed. The strongest result was not that agents replace machine translation, but that a review loop can catch context and fluency issues that a single-pass translation may miss.

View the full publication on SMU Scholar.