okalai-ai/mtig

Grammar-as-control framework for typology-aware synthetic data generation in low-resource languages. # mTIG: Modular Typology-Informed Generation **ACL 2026** mTIG is a framework for **grammar-controlled synthetic data generation** in low-resource languages. It transforms descriptive grammars into executable *control units* that steer large language models toward **typologically balanced** outputs. --- ## 🚀 Overview Large language models (LLMs) can generate synthetic data for low-resource languages, but standard prompting suffers from: - mode collapse (repetitive structures) - poor coverage of morphosyntactic phenomena - unbalanced training distributions mTIG addresses this by introducing **grammar-as-control**: - Decompose grammars into **modular slices** - Use slices as **structured prompts** - Generate data with **explicit distributional control** The result is synthetic corpora that: - cover a wide range of grammatical phenomena - maintain high lexical diversity - improve downstream model performance --- ## 🧠 Core Idea Instead of asking: > *“Can the model generate diverse text?”* mTIG asks: > *“Can we control what the model generates?”* Each **grammar slice** targets a specific phenomenon: - passive voice - causative morphology - noun-class agreement - clause linking - etc. By composing slices, we shape the **distribution of the dataset**, not just individual outputs. --- ## 📊 Key Results - +19% improvement in **typological entropy** - Up to **+20 chrF** in machine translation - Demonstrates a **student-beats-teacher effect**: small MT models trained on mTIG data outperform the source LLM --- ## 🏗️ Pipeline 1. **Grammar decomposition** Convert descriptive grammars into modular slices 2. **Controlled generation** Generate parallel data (e.g., English ↔ target language) 3. **Downstream training** Train MT or other models on the generated corpus --- ## 🌍 Why It Matters Most approaches scale data by **volume**. mTIG scales data by **coverage**. This is especially critical for the *long tail* of languages, where: - large corpora do not exist …