Durak

Durak is a high-performance Turkish NLP toolkit built on a "Rust Core, Python Interface" architecture. Heavy lifting (normalization, tokenization, lemmatization) runs in compiled Rust, releasing the GIL for true parallelism, while providing a flexible, PyTorch-like API for Python researchers.

Why Durak?

• Rust-Powered: Blazing fast text processing with zero-overhead resource embedding
• True Parallelism: GIL-released operations for multi-core batch processing
• Zero-Dependency Distribution: Resources compiled directly into binary
• Research-Ready: Type-safe, reproducible, easy to integrate

Architecture

┌──────────────────────────────────────┐
│   Python Interface (python/durak/)   │  ← Your code here
│   Pipeline • StopwordManager • API   │
├──────────────────────────────────────┤
│   Rust Core (src/lib.rs)             │  ← Performance critical
│   Tokenization • Normalization       │
│   Embedded Resources (include_str!)  │
└──────────────────────────────────────┘

See ARCHITECTURE.md for detailed design documentation.

Quickstart

Installation

pip install durak-nlp

Minimal Pipeline

from durak import process_text

entries = [
    "Türkiye'de NLP zor. Durak kolaylaştırır.",
    "Ankara ' da kaldım.",
]

tokens = [
    process_text(
        entry,
        remove_stopwords=True,
        rejoin_suffixes=True,  # glue detached suffixes before filtering
    )
    for entry in entries
]

print(tokens[0])
# ["türkiye'de", "nlp", "zor", ".", "durak", "kolaylaştırır", "."]

print(tokens[1])
# ["ankara'da", "kaldım", "."]

The pipeline executes: clean → tokenize → rejoin detached suffixes → remove stopwords

Build Blocks à la Carte

from durak import (
    StopwordManager,
    attach_detached_suffixes,
    clean_text,
    remove_stopwords,
    tokenize,
)

text = "İstanbul ' a vapurla geçtik."
cleaned = clean_text(text)
tokens = tokenize(cleaned)
tokens = attach_detached_suffixes(tokens)

# Custom stopword management
manager = StopwordManager(additions=["vapurla"], keep=["istanbul'a"])
filtered = remove_stopwords(tokens, manager=manager)

print(filtered)
# ["istanbul'a", "geçtik", "."]

Accessing the Rust Core

from durak import _durak_core

# High-performance functions (5-10x faster than Python)
normalized = _durak_core.fast_normalize("İSTANBUL")  # "istanbul"
tokens = _durak_core.tokenize_with_offsets("Merhaba dünya!")

# Embedded resources (no file I/O!)
stopwords = _durak_core.get_stopwords_base()  # 100-1000x faster loading
suffixes = _durak_core.get_detached_suffixes()

Features

• Unicode-aware cleaning: Turkish-specific normalization (İ/ı, I/i handling)
• Configurable stopword management: Keep-lists, custom additions, domain-specific sets
• Regex-based tokenizer: Preserves Turkish morphology (clitics, suffixes, apostrophes)
• Offset tracking: Character-accurate positions for NER and span tasks
• Embedded resources: Zero file I/O, compiled directly into binary
• Type-safe: Complete .pyi stubs for IDE support and static analysis
• Tiered lemmatization: Dictionary lookup + heuristic fallback with performance metrics

Lemmatization

Durak provides a high-performance lemmatizer with three strategies:

from durak import Lemmatizer

# Strategy options: "lookup", "heuristic", "hybrid"
lemmatizer = Lemmatizer(strategy="hybrid")

lemmatizer("kitaplar")    # "kitap"
lemmatizer("geliyorum")   # "gel"
lemmatizer("evlerimde")   # "ev"

Strategies

• lookup: Dictionary-only (fastest, high precision, fails on OOV words)
• heuristic: Suffix stripping (handles OOV, may over-strip)
• hybrid: Try lookup first, fallback to heuristic (recommended)

Performance Metrics

Enable metrics to analyze lemmatization behavior:

lemmatizer = Lemmatizer(strategy="hybrid", collect_metrics=True)

for word in large_corpus:
    lemma = lemmatizer(word)

print(lemmatizer.get_metrics())
# Lemmatizer Metrics:
#   Total Calls: 10,000
#   Lookup Hits: 7,234 (72.3%)
#   Heuristic Fallbacks: 2,766
#   Avg Call Time: 0.042ms
#   Lookup Time: 0.038s
#   Heuristic Time: 0.004s

Metrics include:

• Call counts (lookup hits/misses, heuristic fallbacks)
• Timing breakdown (per-strategy latency)
• Cache hit rate
• Average call time

Use cases:

• Compare strategies on your corpus
• Debug lemmatization issues
• Optimize production pipelines
• Report research performance

See examples/lemmatizer_metrics.py for strategy comparison examples.

Root Validation

Control heuristic quality with root validation:

lemmatizer = Lemmatizer(
    strategy="hybrid",
    validate_roots=True,        # Enable validation
    strict_validation=True,     # Require roots in dictionary
    min_root_length=3,          # Minimum 3 characters
)

lemmatizer("kitaplardan")  # Only strips if root ≥3 chars and valid

Development Setup

Building from Source

# Clone the repository
git clone https://github.com/fbkaragoz/durak.git
cd durak

# Create virtual environment
python -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

# Install Rust (if not already installed)
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# Build and install with Rust extension
pip install maturin
maturin develop  # or: maturin develop --release for optimized build

# Install dev dependencies
pip install -e .[dev]

Running Tests

# Run all tests
pytest

# With coverage
pytest --cov=durak --cov-report=html

# Type checking
mypy python

# Linting
ruff check .

Project Structure

durak/
├── src/                  # Rust source (engine)
│   └── lib.rs
├── python/               # Python source (interface)
│   └── durak/
├── resources/            # Static data files
│   └── tr/               # Turkish resources
├── tests/                # Integration tests
├── benchmarks/           # Performance benchmarks
├── examples/             # Usage examples
└── docs/                 # Documentation

Documentation

• Architecture Guide: Design principles and component architecture
• Examples: Basic and advanced usage demonstrations
• Benchmarks: Performance comparison and optimization tips
• API Design Docs: Detailed component specifications
• Changelog: Version history and migration guides
• Roadmap: Future enhancements and planned features

Community & Support

• Code of Conduct: CODE_OF_CONDUCT.md
• Contributing: CONTRIBUTING.md
• Security: SECURITY.md
• Citation: CITATION.cff
• Issues: GitHub Issues

Topics: turkish-nlp, nlp, rust, pyo3, maturin, tokenization, lemmatization, text-processing

Performance

Rust-accelerated functions provide significant speedups:

• Normalization: 5-10x faster than pure Python
• Tokenization: 3-5x faster with offset tracking
• Resource Loading: 100-1000x faster (embedded, no file I/O)
• Full Pipeline: 2-4x overall speedup

Run python benchmarks/benchmark_rust_vs_python.py to measure on your system.

License

Durak is distributed under the Durak License v1.2. Commercial or institutional use requires explicit written permission from the author.

---

Homepage: karagoz.io Repository: github.com/fbkaragoz/durak