Structure-guided generation and verification: constructing high-quality parallel corpora for low-resource agglutinative languages
Journal of King Saud University Computer and Information Sciences · 2026-07 · 38 (5)
DOI: 10.1007/s44443-026-00638-1 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of King Saud University Computer and Information Sciences, online erschienen am 14.03.2026. Verfasst von Yuan Qi und 1 weiteren Autor:innen. Die Arbeit nennt 33 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: National Natural Science Foundation of China, Humanities and Social Sciences Youth Foundation, Ministry of Education.
Abstract (Originalsprache)
While Large Language Models (LLMs) excel at generating fluent text, they frequently fail in low-resource, morphologically complex settings (such as Uzbek, Kazakh, Kyrgyz). In these scenarios, models often prioritize surface-level fluency over structural fidelity, leading to plausible-sounding but factually incorrect hallucinations. To address this, we introduce the Structure-Guided Generation and Verification (SGV) framework, a two-stage approach designed to construct high-quality parallel corpora. First, we use Interlinear Glossed Text (IGT) as a structural constraint, guiding the LLM to follow the grammatical rules of the source language. Second, we employ a Dual-Encoder Verification module to detect and filter out translations that deviate from these constraints. Extensive experiments on three Turkic-Chinese translation tasks demonstrate that SGV significantly outperforms competitive baselines, including LASER and LaBSE mining strategies. Crucially, our analysis reveals a Resilience Gap: the framework maintains high performance even when upstream linguistic tools are noisy, effectively isolating final corpus quality from tool imperfections. Furthermore, we find that SGV is model-agnostic, providing substantial gains across different LLM architectures. These findings suggest that incorporating explicit structural guidance is a more effective strategy for low-resource translation than relying solely on scale or implicit learning. The code is available at https://github.com/almjan/SGV-LowRes-NMT
Förderung
- National Natural Science Foundation of China
- Humanities and Social Sciences Youth Foundation, Ministry of Education
Zitieren
APA
Qi, Y., & Murat, A. (2026). Structure-guided generation and verification: constructing high-quality parallel corpora for low-resource agglutinative languages. Journal of King Saud University Computer and Information Sciences, 38(5). https://doi.org/10.1007/s44443-026-00638-1
Vancouver
Qi Y, Murat A. Structure-guided generation and verification: constructing high-quality parallel corpora for low-resource agglutinative languages. Journal of King Saud University Computer and Information Sciences. 2026;38(5). doi:10.1007/s44443-026-00638-1
BibTeX
@article{Qi2026,
author = {Qi, Yuan and Murat, Alim},
title = {Structure-guided generation and verification: constructing high-quality parallel corpora for low-resource agglutinative languages},
journal = {Journal of King Saud University Computer and Information Sciences},
year = {2026},
volume = {38},
number = {5},
doi = {10.1007/s44443-026-00638-1}
}
EUVIMED