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

2026-04-22 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

A novel in-context learning approach significantly improves low-resource machine translation from Coptic to English by augmenting prompts with syntactic information from Universal Dependencies (UD) parses. Researchers from Georgetown University's Corpling Lab found that while dictionary-based glosses (LEX) are effective, combining them with syntactic data (SYN) yields new state-of-the-art results for Coptic translation across various model sizes, including Gemma and GPT-4.1. The method explores different syntactic representations, such as raw parser outputs (CoNLLU), verbalized dependency relations (DEP), and targeted instructions for difficult constructions (CON). This combined approach consistently outperforms lexicon-only augmentation, demonstrating that even high-quality automatic parses are sufficient and that the benefits extend to out-of-domain texts like Coptic ostraca.

Key takeaway

For research scientists developing machine translation systems for low-resource languages, integrating syntactic information from Universal Dependencies parses alongside traditional dictionary-based augmentation is crucial. This approach, particularly the LEX+SYN combination, has been shown to achieve statistically significant improvements in translation quality for Coptic. You should consider operationalizing syntactic data through raw CoNLL-U outputs, verbalized dependency relations, and targeted instructions for difficult grammatical constructions to maximize performance, even with automatically generated parses.

Key insights

Syntactic augmentation, combined with lexical information, significantly boosts low-resource machine translation quality.

Principles

• Syntactic information complements lexical data for MT.
• High-quality automatic parses are sufficient for ICL MT.
• Model performance varies with syntactic representation.

Method

The method involves augmenting LLM prompts with dictionary entries and various forms of Universal Dependencies syntactic parses, including raw CoNLL-U, verbalized dependencies, and targeted construction-specific instructions.

In practice

• Combine dictionary glosses with UD parses for LRL MT.
• Experiment with raw vs. verbalized syntactic inputs.
• Use construction-specific instructions for complex grammar.

Topics

• Low-Resource Machine Translation
• Coptic Language
• Universal Dependencies
• In-Context Learning
• Syntactic Augmentation

Code references

• gucorpling/in-context-coptic-translation
• mjpost/sacrebleu
• features/copilot

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Syntax as a Rosetta Stone: Universal Dependencies for In-Context Coptic Translation — Syntactic augmentation via Universal Dependencies significantly improves low-resource Coptic-to-English machine translation when combined with dictionary glosses.
• PaliBench: A Multi-Reference Blueprint for Classical Language Translation Benchmarks — Multi-reference benchmarks are essential for evaluating classical language translation due to inherent interpretive plurality.
• AfriSUD: A Dependency Treebank Collection for Evaluating Models on African Languages — AfriSUD provides the first large-scale, native-speaker verified dependency treebanks for nine diverse African languages, revealing a significant syntax gap in current NLP models.
• A UD Parser to the Rescue: A Method for Bringing a Classical Annotated Corpus to Life Again — A knowledge-rich parsing method can effectively update classical linguistic corpora to modern annotation standards.
• Improving Machine Translation of Idioms: A Spanish–Galician Parallel Dataset and Synthetic Augmentation Approach — Synthetic data augmentation using LLMs significantly improves idiom translation in neural machine translation.
• Universal Dependencies v2.5 Benchmarks for spaCy — spaCy v3.2 demonstrates competitive performance against Stanza and Trankit on Universal Dependencies v2.5 benchmarks using CoNLL 2018 evaluation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.