Maedang/MATT_MultiAgentTranslationTeam

Enhancing Low Resource Language Translation in Large Language Models through Multi-agent Workflow # Multi-Agent Translation Team (MATT) Enhancing Low-Resource Language Translation in Large Language Models through a Multi-Agent Workflow --- ## Overview The **Multi-Agent Translation Team (MATT)** project explores how **agentic Large Language Model (LLM) workflows** can improve translation quality for **low-resource languages**. Instead of relying on a single translation model, this system decomposes the translation task into multiple specialized agents that collaborate to iteratively refine outputs. The project evaluates whether structured, multi-agent collaboration can outperform baseline and single-agent translation approaches across multiple quality dimensions. --- ## Motivation & Problem Statement Low-resource languages are often underserved by traditional machine translation systems due to: - Limited parallel training data - Cultural and contextual nuances - Terminology inconsistencies This research investigates whether **multi-agent orchestration**—where each agent focuses on a specific aspect of translation—can: - Improve translation accuracy and fluency - Preserve cultural context and terminology - Produce more consistent and explainable translation outputs --- ## Approach The MATT workflow uses **multiple LLM-based agents**, each assigned a specialized role, such as: - Initial translation - Fluency and grammar refinement - Terminology consistency checking - Cultural and contextual adaptation Agent outputs are combined through a structured workflow to produce a final translation. This agentic approach is compared against: - Baseline translation methods - Single-agent LLM translations --- ## Evaluation Criteria Translations are evaluated using a rubric designed for human-aligned quality assessment, including: - **Accuracy** – Faithfulness to source meaning - **Fluency** – Grammatical correctness and naturalness - **Style** – Tone and readability - **Terminology** – Consistency and domain correctness - **Cultural Context** – Appropriateness f …