uzmorph-nn: Uzbek Neural Morphological Analyzer

uzmorph-nn is a high-accuracy word-level morphological analyzer for the Uzbek language based on a character-level Bidirectional LSTM (BiLSTM) architecture.

Role & Performance

This package provides a robust foundation for Uzbek NLP tasks. It is specifically optimized for:

• Agglutinative Processing: Efficiently decomposes long chains of suffixes.
• Phonological Awareness: Handles stem changes (allomorphy) and vowel/consonant harmony correctly.
• Rule-Augmented Deep Learning: Built on 100,000+ rules from the Common Stem Expansion (CSE) framework.

Installation

pip install uzmorph-nn

Quick Start & Usage Examples

1. Basic Analysis (String Output)

Great for quick debugging or readable logs.

from uzmorph_nn import uzmorph_nn

# Initialize the analyzer
analyzer = uzmorph_nn()

# Analyze a word
result = analyzer.analyze("maktabimizda")
print(result)

# Output:
# Result: 'maktabimizda' -> Stem: maktab | POS: NOUN | Tags: [possession=1, cases=Locative, plural=1]

2. Structured Data (Dictionary)

Ideal for integrating into other projects or data processing pipelines.

result = analyzer.analyze("kitobim")
data = result.to_dict()
print(data)

# Output:
# {
#   "word": "kitobim",
#   "stem": "kitob",
#   "pos": "NOUN",
#   "possession": "1"
# }

3. API Integration (JSON Output)

Useful for web services or sending data between different languages.

result = analyzer.analyze("yozayapmiz")
print(result.to_json())

# Output:
# {
#   "word": "yozayapmiz",
#   "stem": "yoz",
#   "pos": "VERB",
#   "aspect": "Progressive",
#   "person": "1",
#   "number": "Plural"
# }

4. Direct Attribute Access

Access specific parts of the analysis directly.

result = analyzer.analyze("olmalar")

print(f"Stem: {result.stem}")       # Stem: olma
print(f"POS: {result.pos}")         # POS: NOUN
print(f"Features: {result.features}") # Features: ['plural=1']

Architecture Details

• Input: Character sequence (UTF-8).
• Architecture: 2-layer Bidirectional LSTM.
• Tagging Strategy: BIO (Beginning, Inside, Outside) sequence labeling.
• Weights: Pre-trained on a comprehensive CSE rule-engine dataset.

License

MIT License. Free for academic and commercial use.