A Hybrid Rule-Based–Neural Framework with Multi-Criteria Model Selection for Automatic Conversion of Old Uzbek Perso-Arabic Texts into Modern Latin-Script Uzbek
Computer and Decision Making: An International Journal · 07.11.2026 · 3 (8S) · 124-133
DOI: 10.1922/comdem.v3i8s.19141 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer and Decision Making: An International Journal, online erschienen am 07.11.2026. Verfasst von Iskandarova Sayyora Nurmamatovna und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
A substantial part of the written heritage of the Uzbek language — Chagatai (Old Uzbek) manuscripts, lithographs and early printed books produced from the 15th to the early 20th century — exists only in the Perso-Arabic script and therefore remains inaccessible both to modern readers and to Latin-script natural language processing tools. This paper formalizes the automatic conversion of Old Uzbek Perso-Arabic text into modern Latin-script Uzbek as an ill-posed inverse orthographic problem. We prove that the forward orthographic mapping is non-injective and derive a computable upper bound on the token accuracy of any deterministic context-free decoder; instantiated on our test distribution, the corresponding oracle memoryless decoder cannot exceed 35.5% word accuracy, which shows that grapheme-level substitution rules alone are insufficient in principle. Using a fully reproducible synthetic parallel-corpus methodology (967 lexical roots, 7,775 word forms, root-disjoint splits), we compare four systems: a frequency rule-based decoder, a hybrid finite-state candidate generator with an n-gram language model, a BiLSTM encoder–decoder with attention, and a Transformer. Both neural models exceed the oracle context-free reference by about 15 percentage points (50.2% word accuracy; character error rate 8.51–8.89%), whereas the hybrid decoder dominates in the low-data regime. A TOPSIS multi-criteria analysis over accuracy, error rate, latency, model size and data efficiency ranks the hybrid decoder first (CC = 0.763), demonstrating that the accuracy-optimal model is not necessarily the deployment-optimal choice. The results support a two-stage architecture for heritage digitization pipelines and provide a decision-support procedure for selecting a transliteration engine under practical constraints.
Zitieren
APA
Nurmamatovna, I. S., & Begimkulova, P. (2026). A Hybrid Rule-Based–Neural Framework with Multi-Criteria Model Selection for Automatic Conversion of Old Uzbek Perso-Arabic Texts into Modern Latin-Script Uzbek. Computer and Decision Making: An International Journal, 3(8S), 124-133. https://doi.org/10.1922/comdem.v3i8s.19141
Vancouver
Nurmamatovna IS, Begimkulova P. A Hybrid Rule-Based–Neural Framework with Multi-Criteria Model Selection for Automatic Conversion of Old Uzbek Perso-Arabic Texts into Modern Latin-Script Uzbek. Computer and Decision Making: An International Journal. 2026;3(8S):124-133. doi:10.1922/comdem.v3i8s.19141
BibTeX
@article{Nurmamatovna2026,
author = {Nurmamatovna, Iskandarova Sayyora and Begimkulova, Parizod},
title = {A Hybrid Rule-Based–Neural Framework with Multi-Criteria Model Selection for Automatic Conversion of Old Uzbek Perso-Arabic Texts into Modern Latin-Script Uzbek},
journal = {Computer and Decision Making: An International Journal},
year = {2026},
volume = {3},
number = {8S},
pages = {124--133},
doi = {10.1922/comdem.v3i8s.19141}
}
EUVIMED