A Factorial Study of Synthetic Data Generation for Low-Resource Machine Translation using Grammar Books

2026-07-24 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

A new pipeline employs large language models to extract grammatical rules, example sentences, and lexicons directly from descriptive grammar books. This extracted information is then used to generate synthetic parallel corpora, specifically for fine-tuning machine translation models for endangered, low-resource languages. Unlike prior approaches that feed grammar content into prompts at inference time, this method focuses on data generation for model training. Validated across Kalamang (Papuan), Tuatschin (Romance), and Mandan (Siouan), the synthetic data fine-tuning improved over seed-data baselines in 75% of configurations for Kalamang and 59% for Tuatschin, achieving best-case ChrF++ gains of +8.8, +5.3, and +3.3 respectively. A systematic factorial study across 96 configurations identified optimal factor combinations for these gains.

Key takeaway

For NLP Engineers developing machine translation systems for severely under-resourced languages, this research offers a practical, data-centric approach. Instead of relying solely on limited parallel corpora or complex prompting at inference, you should explore using existing grammar books to generate synthetic training data. This method, validated with significant ChrF++ gains, provides a scalable path to build translation tools where traditional data is scarce. Consider the factorial study's insights on part-of-speech, retrieval granularity, and sample volume to optimize your synthetic data generation.

Key insights

Grammar books can be systematically converted into synthetic parallel data for low-resource machine translation fine-tuning.

Principles

• Fine-tuning on synthetic data outperforms inference-time prompting.
• Factorial studies reveal optimal data generation parameters.
• Static linguistic documentation is a valuable MT resource.

Method

A pipeline uses LLMs to extract rules, examples, and lexicons from grammar books, then generates synthetic parallel corpora for machine translation model fine-tuning.

In practice

• Repurpose existing grammar books for MT data creation.
• Experiment with PoS, retrieval granularity, sample volume.
• Prioritize fine-tuning over inference-time prompting for grammar.

Topics

• Low-Resource Machine Translation
• Synthetic Data Generation
• Large Language Models
• Grammar Books
• Model Fine-tuning
• ChrF++

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• A Factorial Study of Synthetic Data Generation for Low-Resource Machine Translation using Grammar Books — Grammar books can be semi-automatically processed to generate synthetic parallel data for low-resource MT fine-tuning.
• Efficient Financial Language Understanding via Distillation with Synthetic Data — Distillation with synthetic data, guided by clustering real examples, enables efficient financial NLP in low-resource settings.
• Data Synthesis and Parameter-Efficient Fine-Tuning for Low-Resource NMT: A Case Study on Q'eqchi' Mayan — Synthetic data effectively bootstraps NMT structural learning for low-resource languages, but requires authentic data for semantic depth.
• An Interactive System for Generating Revisable Grammar Lessons for Extremely Low-Resource Languages Without Expert Annotation — The system enables non-experts to generate revisable grammar lessons for low-resource languages using a human-in-the-loop approach.
• 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.
• I spent 4 days training an AI model to understand my country. It barely works — Low-resource NLP projects face significant challenges in data collection, translation, and accurate labeling, often yielding modest baselines.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.