malaysian-manglish-nlp

Full NLP toolkit for Malaysian Manglish — 51 modules, zero dependencies for core.

Built for real-world Malaysian text: social media, news, chat messages, code-switched Malay-English content.

Installation

pip install malaysian-manglish-nlp

Extras

pip install malaysian-manglish-nlp[transformers]   # HuggingFace transformer models
pip install malaysian-manglish-nlp[embeddings]     # Word2Vec/FastText embeddings
pip install malaysian-manglish-nlp[spacy]          # spaCy integration
pip install malaysian-manglish-nlp[api]            # FastAPI REST API
pip install malaysian-manglish-nlp[langchain]      # LangChain tools
pip install malaysian-manglish-nlp[all]            # Everything

Quick Start

from malaysian_manglish_nlp import sentiment, normalize, ner, detect_language

# Sentiment analysis
result = sentiment("Weh best gila makanan dia!")
print(result)
# {'sentiment': 'positive', 'score': 0.94, 'raw_score': 2.5}

# Text normalization
clean = normalize("xpe la bro, aku ok je")
print(clean)  # "tidak apa la bro, aku okay sahaja"

# Named Entity Recognition
entities = ner("Ali pergi Pavilion KL semalam")
print(entities)
# [{'text': 'Ali', 'type': 'PERSON', 'start': 0, 'end': 3},
#  {'text': 'Pavilion KL', 'type': 'LOCATION', 'start': 10, 'end': 21}]

# Language detection
lang = detect_language("Eh jom makan, I'm hungry gila")
print(lang)
# {'language': 'manglish', 'confidence': 0.87}

HuggingFace

	Link
🤗 Model	vexccz/manglish-nlp-sentiment — XLM-Roberta multi-task (sentiment 95.0%, emotion 90.3%, intent 97.5%)
🤗 Dataset	vexccz/manglish-nlp-dataset — 14,384 labeled Manglish examples
🤗 Demo	vexccz/manglish-nlp-demo — Gradio interactive demo (7 tabs)

Features

Text Processing

• normalize — Expand shortforms (638+ mappings: nk→nak, mcm→macam, sbb→sebab)
• clean — Remove URLs, mentions, repeated chars, HTML
• formalize — Convert informal to formal Malay (aku→saya, ko→anda)
• tokenize — Malaysian-aware tokenizer (handles URLs, hashtags, emoticons)
• stemmer — Malay stemmer with nasal assimilation (250+ roots)
• segment — Sentence segmentation for code-switched text
• spelling — Spell checking with Malaysian dictionary

Analysis

• sentiment — Sentiment analysis with aspect-based (food, service, price, etc.)
• emotion — 8 emotion categories (happy, sad, angry, fear, surprise, disgust, love, neutral)
• sarcasm — Sarcasm detection for Malaysian text
• hate_speech — Hate speech detection (6 categories, severity levels)
• intent — 8 intent types (question, request, complaint, greeting, opinion, statement, command, offer)
• topic — 12 topic classification (food, politics, sports, tech, education, etc.)
• stance — Stance detection (support/oppose/neutral)
• profanity — Profanity detection with leetspeak evasion handling

Entity & Structure

• ner — Named Entity Recognition (11 types: PERSON, ORG, LOC, PRODUCT, EVENT, MONEY, PHONE, EMAIL, DATE, TIME, PERCENT)
• pos_tag — Part-of-speech tagging (15 tags)
• dependency — Dependency parsing (SVO extraction)
• coreference — Pronoun resolution with Malaysian gender heuristics
• keywords — Keyword extraction (frequency, RAKE, TF-IDF, TextRank)

Language Detection & Code-Switching

• language — Language identification (Malay/English/Manglish/Mixed)
• code_switching — Code-switching point detection, switch ratio, segmentation by language
• dialect — 6 Malay dialects (Standard, Kelantan, Terengganu, N9, Kedah, Sarawak, Sabah) with normalization

Semantic & Similarity

• similarity — Text similarity (Jaccard, cosine, overlap, semantic)
• embeddings — Word2Vec/FastText trained on Malaysian social media (518 vocab, 100d)
• augmentation — Text augmentation (synonym replacement, shortform variation)

Generation & Understanding

• translation — Rule-based BM↔EN translation (1000+ word pairs, phrase translation)
• summarization — Extractive summarization using TextRank algorithm
• text_generation — N-gram based text generation and autocomplete
• qa — Extractive question answering with TF-IDF retrieval
• discourse — Argument mining and fallacy detection
• ocr_normalize — OCR text correction for Malaysian documents

Preprocessing & Utilities

• normalizer — Advanced normalization (money, dates, times, elongated text)
• dictionary — Malay-English dictionary lookup
• similarity — Multiple similarity metrics
• pipeline — Chain multiple modules together
• calibration — Confidence scoring for predictions
• hybrid_ml — Feature extraction + logistic classifier
• evaluate — Model evaluation and regression tracking
• cache — LRU caching for performance
• profiler — Performance benchmarking tools
• tuning — Hyperparameter tuning and threshold optimization

Integration

• spacy_integration — Custom spaCy Language class and pipeline components
• rest_api — FastAPI REST API with rate limiting and CORS
• langchain_tool — LangChain tool wrappers
• CLI — Command-line interface with subcommands

Fine-Tuned Model

An XLM-Roberta multi-task model fine-tuned on 14,384 labeled Manglish examples for sentiment, emotion, and intent classification.

from malaysian_manglish_nlp.transformers.manglish_model import load_model, predict

model = load_model()  # Auto-downloads from HuggingFace on first use
result = predict("gila best servis ni")
# {'sentiment': {'label': 'positive', 'confidence': 0.96},
#  'emotion':    {'label': 'happy',    'confidence': 0.85},
#  'intent':     {'label': 'opinion',  'confidence': 1.00}}

Task	Accuracy	Classes
Sentiment	95.0%	positive, negative, neutral
Emotion	90.3%	happy, sad, angry, fear, surprise, disgust, love, neutral
Intent	97.5%	question, statement, request, complaint, greeting, opinion, command, offer
Average	94.3%	—

Model: vexccz/manglish-nlp-sentiment on HuggingFace. Requires pip install malaysian-manglish-nlp[transformers].

Performance

• 23,000+ texts/sec sentiment analysis throughput (rule-based)
• <0.5s import time for core modules
• Zero dependencies for core text processing
• LRU caching on heavy operations (stemmer, normalize, sentiment, language detection)
• Lazy loading — only imports what you use
• Pre-compiled regex patterns across 6 modules

Comparison with Malaya

Feature	malaysian-manglish-nlp	Malaya
Core dependencies	None	TensorFlow/PyTorch required
Import time	<0.5s	10-30s
Manglish-first	Built for informal MY text	Formal BM focus
Modules	51	~40
Throughput	23k+ texts/sec	Varies (GPU recommended)
Python support	3.8-3.12	3.8+
Aspect sentiment	✅	❌
Code-switching detection	✅	❌
Hate speech detection	✅	Limited
Discourse analysis	✅	❌
OCR normalization	✅	❌
Translation (rule-based)	✅	❌
Text generation	✅	❌

Both are solid choices. Malaya excels at formal Bahasa Melayu with deep learning models. malaysian-manglish-nlp is optimized for informal, code-switched Malaysian text with minimal overhead and advanced NLP features.

CLI Usage

# Full analysis
manglish analyze "Weh best gila makanan dia!"

# Sentiment
manglish sentiment "Teruk la service kat sini"

# Normalize shortforms
manglish normalize "xpe la bro aku otw"

# Translate
manglish translate "Aku nak pergi makan" --to en

# NER
manglish ner "Ahmad kerja kat Google Malaysia"

# Summarize file
manglish summarize --file article.txt

# Run benchmarks
manglish benchmark

# Profile performance
manglish profile "Sample text here"

REST API

# Start API server
uvicorn malaysian_manglish_nlp.rest_api:app --port 8000

# Or with Docker
docker-compose up -d

Endpoints:

• POST /analyze — Full analysis
• POST /sentiment — Sentiment only
• POST /normalize — Normalize text
• POST /translate — Translate text
• POST /ner — Named entities
• POST /pos — POS tags
• POST /summarize — Summarize text
• POST /batch — Batch process multiple texts
• GET /health — Health check
• GET /modules — List available modules

Testing

# Run all tests (900+ tests)
python -m pytest tests/ -q

# Run specific test file
python -m pytest tests/test_sentiment.py -v

# Run with coverage
python -m pytest tests/ --cov=malaysian_manglish_nlp

# Run heavy tests (requires gensim)
RUN_HEAVY_TESTS=1 python -m pytest tests/test_word_embeddings.py -v

Documentation

Full documentation available at manglish-nlp.readthedocs.io

Includes:

• Module reference for all 51 modules
• API documentation with examples
• Performance benchmarks
• Comparison with Malaya
• Contributing guide
• Changelog

Contributing

Contributions welcome! Areas where help is needed:

1. More training data — Manglish text samples from social media
2. Dialect support — More regional variants and normalization rules
3. Benchmarks — Comparative benchmarks on Malaysian NLP datasets
4. Documentation — More usage examples and tutorials

git clone https://github.com/ZafranYusof/malaysia-manglish-nlp.git
cd malaysian-manglish-nlp
pip install -e ".[all]"
python -m pytest tests/ -q

License

MIT — see LICENSE for details.

Citation

If you use malaysian-manglish-nlp in your research, please cite:

@software{malaysian_manglish_nlp,
  author = {Zafran},
  title = {malaysian-manglish-nlp: Full NLP toolkit for Malaysian Manglish},
  year = {2026},
  url = {https://github.com/ZafranYusof/malaysia-manglish-nlp},
  version = {3.2.0}
}