GlossMATE: Multi-Agent Translator Explanations for Glosses

Abstract

This paper introduces GlossMATE, a multi-agent critique-and-judge system that translates the gloss line in Interlinear Glossed Text (IGT) into fluent English using Large Language Models (LLMs). GlossMATE integrates linguist-provided resources (e.g., gloss-tag explanations, lexicon entries, curated IGT) with in-context learning and a multi-agent critique-and-judge procedure that iteratively evaluates and refines candidate translations. Our experiments show that leveraging analogous examples, explicit linguistic explanations, and collaborative agent interactions can enhance translation quality across several low-resource and polysynthetic languages. We also incorporate human linguists into the critique loop for selected languages. Case studies on three Indigenous languages further demonstrate the complementary strengths of human-in-the-loop feedback and multi-agent reasoning for language documentation tasks.

Resources

Details