basa 0.1.0a1

Modern NLP for Indonesian and regional languages

BASA

Modern NLP preprocessing for Indonesian and regional languages.

BASA is a lightweight, zero-dependency preprocessing library designed for real-world Indonesian text — the kind found on Twitter/X, TikTok, WhatsApp, Shopee reviews, and Discord. It normalizes informal slang, collapses expressive character repetition, reduces punctuation noise, and optionally corrects typos, all through a single clean API.

from basa import normalize

normalize("GW GKKKK NGERTIII BNGTTTT!!!!!")
# → 'saya tidak mengerti banget!'

---

Why BASA?

Indonesian social media text is notoriously difficult to process with standard NLP tools:

Raw input	After normalize()
gw gk ngerti bngt sihhhh!!!	saya tidak mengerti banget sih!
kmrn gamau makan krn baper bgt	kemarin tidak mau makan karena bawa perasaan banget
otw gan, rekber dlu ya!!!!!	dalam perjalanan saudara, rekening bersama dulu ya!
GW GKKKK NGERTIII BNGTTTT!!!!!	saya tidak mengerti banget!

Standard tokenizers and language models often fail on this kind of input because they see "gkkkk", "bngtttt", and "ngertiii" as unknown tokens. BASA normalizes them first.

---

Installation

pip install basa

Requires Python 3.10+

---

Quick Start

One-liner (recommended for most use cases)

from basa import normalize

normalize("gw gk ngerti bngt sihhhh!!!")
# → 'saya tidak mengerti banget sih!'

Zero-config alias

from basa import quick

quick("GW GKKKK NGERTIII BNGTTTT!!!!!")
# → 'saya tidak mengerti banget!'

quick() is a thin alias for normalize() with all defaults applied. Use it when you want the shortest possible call.

Batch processing

from basa import normalize

texts = [
    "gw gk ngerti",
    "lu udh makan??",
    "kmrn gamau pergi krn baper bgt",
]

normalize(texts)
# → ['saya tidak mengerti', 'kamu sudah makan?', 'kemarin tidak mau pergi karena bawa perasaan banget']

---

API Reference

normalize(text, **options)

Normalize informal Indonesian text. Accepts a single string or a list of strings.

normalize(
    text: Union[str, List[str]],
    apply_slang: bool = True,
    apply_typo: bool = False,
    lowercase: bool = True,
    normalize_punctuation: bool = True,
    normalize_whitespace: bool = True,
) -> Union[str, List[str]]

Parameters

Parameter	Type	Default	Description
text	str or List[str]	—	Input text or list of texts.
apply_slang	bool	True	Expand slang and reduce expressive repeated characters (e.g. "bngtttt" → "banget").
apply_typo	bool	False	Correct misspelled words using Levenshtein distance. Opt-in — requires a vocabulary to be loaded first.
lowercase	bool	True	Lowercase the text before processing. Set False for NER and case-sensitive pipelines.
normalize_punctuation	bool	True	Collapse repeated punctuation marks ("!!!!!" → "!").
normalize_whitespace	bool	True	Strip leading/trailing whitespace and collapse internal multiple spaces.

Processing pipeline (in order)

1. lowercase            → "GW GK NGERTI" → "gw gk ngerti"
2. slang normalization  → "gkkkk" → "gk" → "tidak"
3. typo correction      → "mkan" → "makan"  (opt-in)
4. punctuation          → "!!!!!" → "!"
5. whitespace cleanup   → "  a   b  " → "a b"

Examples

# Preserve case for NER tasks
normalize("Jokowi pergi ke Jakarta", lowercase=False)
# → 'Jokowi pergi ke Jakarta'

# Disable slang (pass through raw tokens)
normalize("gw gk ngerti", apply_slang=False)
# → 'gw gk ngerti'

# Enable typo correction (requires vocab)
from basa import typo
typo.add_to_vocab({"makan", "minum", "pergi"})
normalize("saya mkan dan mnum", apply_typo=True)
# → 'saya makan dan minum'

---

quick(text)

Zero-config alias for normalize() with all default settings.

from basa import quick

quick("gw gamau pergi krn mager")
# → 'saya tidak mau pergi karena malas bergerak'

---

typo — Typo Corrector

BASA's typo corrector is vocabulary-driven and opt-in by default. You supply the vocabulary; BASA finds the closest match using Levenshtein distance.

from basa import typo

# Load your domain vocabulary
typo.add_to_vocab({"makan", "minum", "masak", "pergi", "datang"})

typo.correct("mkan")     # → 'makan'
typo.correct("mnm")      # → 'minum'
typo.correct("ok")       # → 'ok'  (too short, skipped by default)

# Correct a full sentence
typo.correct_text("saya mkan dan mnm")
# → 'saya makan dan minum'

# Get multiple suggestions
typo.suggest("mkan", top_k=3)
# → ['makan', 'masak', 'minum']

Why is apply_typo=False by default?

Typo correction is destructive when applied blindly. Without the right vocabulary, domain-specific terms like xgboost, lightgbm, or rekber would be mangled. BASA follows the principle of conservative by default, destructive features opt-in.

Vocabulary management

from basa import typo

typo.add_to_vocab({"kata", "lain"})      # add words
typo.remove_from_vocab({"kata"})         # remove words
typo.clear_vocab()                       # reset entirely
len(typo)                                # vocab size
"makan" in typo                          # membership check

# Check cache statistics (useful for profiling)
typo.cache_info()
# → {'hits': 120, 'misses': 35, 'size': 35}

Typo corrector options

from basa.core.typo import TypoCorrector

corrector = TypoCorrector(
    vocab={"makan", "minum"},
    min_word_length=4,    # tokens shorter than this are skipped (default: 4)
    min_confidence=0.5,   # minimum correction confidence in [0, 1] (default: 0.5)
)

---

slang — Slang Normalizer

Access the underlying slang engine directly for fine-grained control.

from basa.core.slang import slang, SlangNormalizer

# Use the singleton
slang.normalize("gw gamau pergi krn lg baper bgt")
# → 'saya tidak mau pergi karena sedang bawa perasaan banget'

# Custom dictionary (extend or override defaults)
custom = SlangNormalizer(custom_mapping={
    "gaskeun": "ayo lakukan",
    "jancok":  "ekspresi",
})
custom.normalize("gaskeun bro!")
# → 'ayo lakukan bro!'

# Batch normalize
slang.normalize_batch(["gw makan", "lu minum"])
# → ['saya makan', 'kamu minum']

Slang dictionary categories

The built-in dictionary covers 250+ entries across 13 categories:

Category	Examples
Pronouns	gw → saya, lu → kamu, dy → dia
Kinship & address	kk → kakak, klg → keluarga, ortu → orang tua
Negation	ga, gak, nggak → tidak
Compound negation	gamau → tidak mau, gabisa → tidak bisa
Conjunctions	yg → yang, krn → karena, tp → tapi
Verbs	udah → sudah, blm → belum, ngerti → mengerti
Adjectives & adverbs	bgt → banget, bener → benar, dikit → sedikit
Question words	gmn → bagaimana, knp → kenapa, kmn → kemana
Greetings & responses	makasih → terima kasih, sip → baik
Temporal & location	skrg → sekarang, kmrn → kemarin, ntr → nanti
Internet slang	otw → dalam perjalanan, btw → omong-omong, wkwk → tertawa
E-commerce & finance	ongkir → ongkos kirim, rekber → rekening bersama, cod → bayar di tempat
Youth / Gen-Z	mager → malas bergerak, baper → bawa perasaan, gabut → tidak ada kegiatan

---

Real-World Use Cases

Preprocessing for sentiment analysis

from basa import normalize

reviews = [
    "produknya bagus bgt tp ongkirnya mahal bgt!!!",
    "gw kecewa bngt, barang ga sesuai deskripsi smskali",
    "rekber dlu gan, takut kena tipu",
]

clean = normalize(reviews)
# Pass clean into your sentiment model

Preprocessing for a custom NLP pipeline

from basa import normalize, typo

# Load your domain vocabulary (e.g., from a word list file)
with open("vocab.txt") as f:
    domain_vocab = set(f.read().splitlines())

typo.add_to_vocab(domain_vocab)

def preprocess(text: str) -> str:
    return normalize(text, apply_typo=True)

preprocess("gw mkan siang tdi di wrng padang")
# → 'saya makan siang tadi di warung padang'

NER pipeline (preserve casing)

from basa import normalize

text = "Jokowi blg bhw pemerintah akan bantu UMKM"
normalize(text, lowercase=False)
# → 'Jokowi bilang bahwa pemerintah akan bantu UMKM'

---

Design Philosophy

BASA is built around three principles:

1. Conservative by default. Only safe, lossless transforms are enabled out of the box. Destructive features (like typo correction) require explicit opt-in.

2. No bundled vocabularies for correction. Every domain has different vocabulary needs — fintech, e-commerce, ML, healthcare. Callers supply their own word list via typo.add_to_vocab().

3. Zero required dependencies for core preprocessing. The normalize() and slang modules use only the Python standard library. The optional transformers, torch, and pydantic dependencies are only required for advanced modules (basa.translate, basa.evaluate).

---

Development

Setup

git clone https://github.com/Muanai/basa.git
cd basa
python -m venv .venv
.venv\Scripts\activate       # Windows
# source .venv/bin/activate  # macOS / Linux
pip install -e ".[dev]"

Running tests

pytest tests/ -v

Optional extras

pip install -e ".[serving]"     # FastAPI serving
pip install -e ".[evaluation]"  # ROUGE, BERTScore, seqeval
pip install -e ".[dev]"         # pytest, ruff, black, mypy

---

Roadmap

Version	Status	Features
v0.1	✅ Current	normalize(), quick(), slang (250+ entries), typo corrector
v0.2	🔜 Planned	BK-Tree / SymSpell for faster typo correction at large vocab sizes
v0.3	🔜 Planned	Emoji handling, remove_emoji flag
v0.4	🔜 Planned	Tokenizer module (basa.tokenize)
v1.0	🔜 Planned	Stable API, full docs site, PyPI release

---

Contributing

Contributions are welcome! In particular:

• Slang dictionary additions — if you spot a common slang word that's missing, open a PR adding it to the appropriate category in src/basa/core/slang.py.
• Bug reports — please include the exact input string and the unexpected output.
• Performance improvements — especially for the typo correction module.

Please open an issue before submitting large changes.

---

License

MIT © 2026 Muanai Khalifah Revindo