Get started
Work

A Systematic Benchmark of Machine Transliteration Models for the Tajik-Farsi Language Pair: A Comparative Study from Rule-Based to Transformer Architectures

Year: 2026
Type: conference-paper
Abstract: This paper presents a systematic comparative benchmark of machine learning architectures for transliteration between the Tajik (Cyrillic script) and Persian (Arabic script) languages. A central contribution is the creation and open release of a unique, multi-domain parallel corpus comprising 328,253 sentence pairs aggregated from heterogeneous sources, including classical poetry, diplomatic texts, lexicographic data, and crowdsourced projects. To enable fully reproducible experiments under limited computational resources, a representative stratified subset of 40,000 pairs has been formed and is made publicly available together with the complete dataset. The study evaluates six model classes under a unified protocol: a rule-based baseline, an LSTM with attention, a character-level Transformer, a G2P Transformer trained from scratch, large pre-trained multilingual models (mBART, mT5 with LoRA), and the byte-oriented ByT5 model. The experimental results demonstrate that ByT5 achieves state-of-the-art performance with chrF++ scores of 87.4 for Tajik→Farsi and 80.1 for the reverse direction. Notably, the proposed G2P Transformer, trained exclusively on the 40,000-pair subset, significantly outperforms the substantially larger mBART model in the forward direction (72.3 vs. 62.2 chrF++), while subword-based models (mT5) fail almost completely, underscoring the necessity of character- or byte-level representations for this task. All collected data, source code, and model outputs are released under open licenses, thereby establishing a FAIR (Findable, Accessible, Interoperable, Reusable) benchmark for cross-script NLP in low-resource Central Asian languages.
DOI: 10.47810/clib.26.13
Language: English
Open Access status: bronze
OpenAlex ID: W7160247692
FWCI: 3.967
Cites: 2
Cited by: 3
Related to: 10