Data Synthesis and Parameter-Efficient Fine-Tuning for Low-Resource NMT: A Case Study on Q'eqchi' Mayan

Researchers developed a data synthesis methodology for neural machine translation of Q'eqchi' Mayan, using synthetic corpora derived from community dictionaries and Parameter-Efficient Fine-Tuning to avoid extractive web-scraping. While the approach achieved strong structural performance (BLEU 42.02 on synthetic data), it revealed a critical gap: the model excels at learning grammar but fails to acquire authentic semantic grounding (BLEU 0.59 on organic text), suggesting synthetic bootstrapping alone cannot replace real-world linguistic diversity.

This research addresses a genuine tension in low-resource language technology: how to build functional NMT systems while respecting data sovereignty and avoiding exploitative data practices. The Q'eqchi' case study reveals both the promise and limitations of synthetic data approaches. By converting community dictionaries into constrained training templates, researchers achieved measurable success in teaching complex morphosyntactic features—agglutination and VOS word order—demonstrating that synthetic constraints can effectively encode linguistic structure.

The critical finding emerges in the performance gap between synthetic and organic evaluation. The model learned rigid structural patterns but couldn't generalize to natural language variation, suggesting it memorized template distributions rather than acquiring flexible linguistic competence. This overfitting problem intensified when multi-task learning was introduced; additional auxiliary tasks competed for the limited capacity within LoRA adapters, causing negative transfer and prioritizing synthetic markers over organic adaptability.

For the broader language technology landscape, this study has significant implications. It establishes synthetic data's utility as a structural primer while demonstrating its insufficiency as a standalone solution. The findings challenge the assumption that parameter-efficient methods alone solve low-resource scenarios; architectural constraints combined with limited authentic data create bottlenecks that fine-tuning cannot overcome. The research highlights that semantic grounding—understanding word meanings in context—requires exposure to genuine linguistic variation that no synthetic pipeline can fully replicate.

Moving forward, practitioners should view synthetic bootstrapping as a stepping stone requiring curriculum-based refinement with authentic data. This framework prioritizes Indigenous language sovereignty while establishing realistic expectations about what synthetic approaches can deliver, pointing toward hybrid methodologies that balance ethical data practices with linguistic authenticity.

• →Synthetic data from dictionaries effectively teaches grammatical structure but fails to provide semantic grounding needed for natural language fluency
• →Parameter-efficient fine-tuning alone cannot overcome the structural-semantic gap created by constrained synthetic training templates
• →Multi-task learning within capacity-limited LoRA adapters caused negative transfer, suggesting architectural trade-offs in low-resource settings
• →Data sovereignty benefits of avoiding web-scraping must be paired with authentic corpus collection for semantic refinement
• →Synthetic bootstrapping works best as a primer requiring curriculum learning with real linguistic data rather than as a standalone solution