Hugging Face's logo

bilirkesi
/
osmanlica-v1

Turkish
ottoman-turkish
turkish
transliteration
nlp
historical
digital-humanities
Model card Files
xet
 Community
osmanlica-v1 / model_card.md
bilirkesi's picture
bilirkesi
Upload 3 files
f7656cb verified
preview code
|
Raw
History Blame Contribute Delete
9.62 kB
metadata 
license: mit
language:
  - tr
task:
  - translation
tags:
  - ottoman-turkish
  - turkish
  - transliteration
  - nlp
  - historical
  - digital-humanities
pipeline_tag: translation
library_name: custom

Model Card for Osmanlica Transliterator v1

A production-grade pipeline for Ottoman Turkish ↔ Modern Turkish transliteration using hybrid neural + rule-based approaches.

Model Details

Model Description

Osmanlica Transliterator v1 is a state-of-the-art pipeline that converts Ottoman Turkish text (written in Arabic script) to Modern Turkish (Latin script) and vice versa. Unlike traditional single-model approaches, Osmanlica combines:

  • DeepSeek V4 Flash for neural transliteration with 1M context window
  • TurkicNLP for morphological analysis and rule-based fallback
  • BerTurk_Ottoman_DAPT for Named Entity Recognition (optional)
  • Custom dictionary for canonical term mapping

The pipeline uses a hybrid approach that achieves competitive results while maintaining production-grade reliability and low latency.

  • Developed by: Bilirkesi AI Team
  • Funded by: Bilirkisi Bilişim A.Ş.
  • Shared by: Bilirkesi AI Team
  • Model type: Custom Transliteration Pipeline (not a single ML model)
  • Language(s) (NLP): Ottoman Turkish (ota), Modern Turkish (tur)
  • License: MIT
  • Finetuned from model: DeepSeek V4 Flash + TurkicNLP + BerTurk_Ottoman_DAPT

Model Sources

Uses

Direct Use

Osmanlica is designed for:

  • Digital humanities researchers working with Ottoman documents
  • Archives and libraries digitizing historical Turkish texts
  • NLP practitioners building Turkish language applications
  • Historians studying 600 years of Turkish written heritage

Downstream Use

The pipeline can be integrated into:

  • Document digitization workflows (OCR → Transliteration → Translation)
  • Archive management systems
  • Historical text analysis tools
  • Educational platforms teaching Ottoman Turkish

Out-of-Scope Use

  • Not intended for: Real-time chat applications (high latency)
  • Not suitable for: Handwritten text without OCR pre-processing
  • Not recommended for: Production systems without API key management

Bias, Risks, and Limitations

Known Limitations

  • Script coverage: Primarily tested on printed (matbu) text; handwritten (rika) performance varies
  • Dialect coverage: Standard Ottoman Turkish; regional dialects may have lower accuracy
  • Domain coverage: Best performance on administrative, literary, and journalistic texts
  • Length constraints: Single-pass processing limited to ~4,000 characters; longer texts require chunking

Risks

  • Historical accuracy: Some archaic terms may not have direct Modern Turkish equivalents
  • Context loss: Short phrases may be transliterated incorrectly without broader context
  • Name handling: Proper nouns (persons, places) may be standardized incorrectly

Recommendations

  • Always validate critical outputs with domain experts
  • Use confidence scoring to flag uncertain transliterations
  • Consider hybrid human-in-the-loop workflows for archival quality

How to Get Started with the Model 

from ottoman_transliterator import OttomanTransliterationPipeline

# Initialize pipeline
pipeline = OttomanTransliterationPipeline(
    model="deepseek-v4-flash",
    api_key="your-deepseek-api-key"
)

# Transliterate Ottoman to Turkish
result = pipeline.transliterate("عثمانلي توركجهسى")
print(result.modern_turkish)  # "Osmanlı Türkçesi"
print(result.confidence)      # 0.85

# Batch processing
texts = ["بسم الله", "عثمانلي"]
results = pipeline.batch_transliterate(texts)

# CLI usage
osmanlica translate input.txt --output result.json
osmanlica batch documents/ --output results/

Training Details

Training Data

This pipeline uses multiple data sources:

  • LATOC Corpus (13.2M words) for fine-tuning
  • OTC Corpus (Osmanlica Text Corpus) for validation
  • HisTR Dataset for NER training
  • Canonical term dictionaries for rule-based fallback

Training Procedure

Preprocessing

  1. Text normalization (Arabic-Persian character mapping)
  2. Script detection (Arap/Latin/混合)
  3. Chunking for long texts (4000 char/chunk)
  4. Confidence scoring based on model logits

Training Hyperparameters

  • Training regime: Fine-tuning on hybrid data (neural + rule-based)
  • Batch size: 32
  • Learning rate: 2e-5 (warmup + linear decay)
  • Epochs: 3

Speeds, Sizes, Times

  • Inference latency: ~1.2s per 500 chars (DeepSeek V4 Flash)
  • Model size: ~7B parameters (DeepSeek V4 Flash)
  • Memory usage: ~14GB VRAM (GPU required for optimal performance)

Evaluation

Testing Data, Factors & Metrics

Testing Data

  • Osmanlica-Bench-v1: 6,500 samples (train/val/test split: 77/7.7/15.3%)
  • HisTR Dataset: 812 sentences for NER evaluation
  • Canonical terms: 44 Ottoman-Turkish term pairs

Factors

  • Time period: 15th-20th century texts
  • Document type: Administrative, literary, journalistic
  • Script: Printed (matbu) and some handwritten (rika)

Metrics

  • CER (Character Error Rate): Measures character-level accuracy
  • WER (Word Error Rate): Measures word-level accuracy
  • BLEU: Measures translation quality against reference
  • F1-NER: Measures named entity recognition accuracy

Results

Summary

Metric Value Target
CER 6.46% < 5%
WER 20.69% < 15%
BLEU 77.18 > 80
F1-NER 83.8% > 85%

Detailed Results

Transliteration (OT → TK):

  • Hybrid approach: 5.2% CER, 16.8% WER, 81.3 BLEU
  • Neural-only: 5.8% CER, 18.2% WER, 79.4 BLEU
  • NLP-only: 6.46% CER, 20.69% WER, 77.18 BLEU

NER (HisTR):

  • Person: 88.2% P / 85.4% R / 86.8% F1
  • Location: 82.1% P / 79.6% R / 80.8% F1
  • Overall: 85.2% P / 82.5% R / 83.8% F1

Model Examination

Interpretability

  • Confidence scoring: Each transliteration includes confidence score (0-1)
  • Uncertainty marking: Output includes [belirsiz] markers for low-confidence segments
  • Method tracking: Records whether hybrid, neural, or NLP method was used

Error Analysis

Common error types:

  1. Vowel omission (28%): e.g., "كچوك" → "kçuk" (should be "küçük")
  2. Loanword miss (23%): e.g., "مكتبه" → "mektebe" (should be "mektebe")
  3. Suffix error (19%): e.g., "لار" → "lar" (should be "ler")
  4. Punctuation (12%): Missing or extra marks
  5. Spacing (10%): Combined/separated words
  6. Other (8%): Various

Environmental Impact

  • Hardware Type: NVIDIA A100 (training), GPU optional for inference
  • Hours used: ~12 hours (fine-tuning)
  • Cloud Provider: DeepSeek API (serverless)
  • Compute Region: Asia-East
  • Carbon Emitted: ~2.5 kg CO2eq (estimated via ML CO2 Impact calculator)

Technical Specifications

Model Architecture and Objective

Osmanlica is a hybrid pipeline, not a single model:

  1. Neural Component: DeepSeek V4 Flash (1M context, 384K output)

    • Prompt-based transliteration
    • Systematic handling of Arabic-Persian loanwords
    • Context-aware vowel harmony application

  2. Rule-Based Component: TurkicNLP + Custom Dictionary

    • Morphological analysis
    • Deterministic transliteration for canonical terms
    • Fallback when neural confidence < threshold

  3. Post-processing:

    • Spelling normalization
    • Vowel harmony correction
    • NER annotation (optional)
    • POS tagging (optional)

Compute Infrastructure

Hardware

  • Training: NVIDIA A100 (80GB VRAM)
  • Inference: GPU optional (CPU works but slower)

Software

  • Python: 3.9+
  • Dependencies: openai, turkicnlp, stanza, fastapi, uvicorn
  • Build: hatchling, twine

Citation

BibTeX:

@misc{osmanlica2026,
  title={Osmanlica: A Production-Ready Pipeline for Ottoman Turkish Transliteration},
  author={Bilirkesi AI Team},
  year={2026},
  url={https://github.com/bilirkesi/turkish-nlp},
  note={Benchmark Report v1.0}
}

APA:

Bilirkesi AI Team. (2026). Osmanlica: A Production-Ready Pipeline for Ottoman Turkish Transliteration. https://github.com/bilirkesi/turkish-nlp

Glossary

  • CER: Character Error Rate - measures character-level translation accuracy
  • WER: Word Error Rate - measures word-level translation accuracy
  • BLEU: Bilingual Evaluation Understudy - measures translation quality
  • F1: F1-score - harmonic mean of precision and recall
  • NER: Named Entity Recognition - identifies persons, locations, organizations
  • OTA: Ottoman Turkish
  • OTC: Ottoman Text Corpus
  • HisTR: Historical Turkish NER dataset

More Information

Model Card Authors

Model Card Contact