Syntax as a Rosetta Stone: Universal Dependencies for In-Context Coptic Translation

A team of computational linguists has published a breakthrough paper, 'Syntax as a Rosetta Stone: Universal Dependencies for In-Context Coptic Translation,' presenting a novel method to significantly improve machine translation for low-resource languages. The research, led by Abhishek Purushothama with co-authors Emma Thronson, Alexia Guo, and Amir Zeldes, addresses the core challenge of translating Coptic, an ancient Egyptian language with limited parallel text data. Instead of traditional model training, their approach uses in-context learning with large language models (LLMs), cleverly augmenting the input prompts with structured linguistic information.

The key innovation is combining two types of data within the prompt: bilingual dictionary glosses for vocabulary and, crucially, syntactic parses from Universal Dependencies. The team experimented with feeding raw parser outputs, verbalized descriptions of sentence structure in plain English, and targeted instructions for difficult grammatical constructions identified in parse sub-trees. While syntactic data alone was less effective than dictionary data, the fusion of both created a powerful 'Rosetta Stone' effect. This hybrid prompt engineering led to substantial performance gains across LLMs of different sizes, ultimately achieving new state-of-the-art translation results for Coptic, demonstrating that explicitly providing grammatical roadmaps can dramatically enhance an LLM's ability to navigate an unfamiliar language.