Skip to content

Typo Correction

BASA provides conservative typo correction based on Levenshtein distance.

Unlike many normalization systems, typo correction is disabled by default.

from basa import normalize

normalize(
    "gw mkan",
    apply_typo=False
)

Output:

"saya mkan"

Why Is Typo Correction Opt-In?

Automatic typo correction can damage:

  • Proper nouns
  • Technical terms
  • Acronyms
  • Domain-specific vocabulary

For example:

xgboost
lightgbm
tensorflow
jokowi
telkomsel

To avoid destructive behavior, BASA requires explicit user consent before applying typo correction.

Loading a Vocabulary

Before enabling typo correction, add words to the vocabulary.

from basa import typo

typo.add_to_vocab({
    "makan",
    "minum",
    "berangkat"
})

Correcting Individual Words

typo.correct("mkan")

Output:

"makan"

Correcting Full Sentences

from basa import normalize

normalize(
    "gw mkan dulu",
    apply_typo=True
)

Output:

"saya makan dulu"

Custom Maximum Distance

The typo engine uses Levenshtein distance.

By default, the maximum correction distance is conservative.

You can also use the lower-level API directly:

typo.correct(
    "mkan",
    max_dist=2
)

Protected Vocabulary

BASA automatically protects normalized slang outputs from being modified by the typo corrector.

For example:

"gk"
→ "tidak"

The word "tidak" will never be reinterpreted as a typo candidate.

This prevents interactions between the slang and typo pipelines from producing incorrect results.

For most use cases, quick() is the recommended entry point:

from basa import quick

quick("gw gk ngerti bngtttt")

This approach provides safe, zero-configuration normalization for informal Indonesian text.

For domain-specific datasets, load your own vocabulary before enabling typo correction:

from basa import normalize, typo

typo.add_to_vocab({
    "xgboost",
    "lightgbm",
    "tensorflow",
    "fintech",
    "telkomsel",
})

normalize(
    "gw belajar xgboosst",
    apply_typo=True,
)

Adding domain-specific terms helps reduce unintended corrections and gives you more control over the normalization pipeline.

Best Practices

When using typo correction, we recommend:

  • Load a vocabulary that matches your domain.
  • Keep apply_typo=False unless spelling correction is required.
  • Add technical terms, company names, and acronyms to the vocabulary.
  • Review normalization outputs before using them in production workflows.

BASA intentionally favors conservative behavior over aggressive correction.

Future Directions

Potential areas for future development include:

  • Frequency-based candidate ranking
  • Optional domain-specific dictionaries
  • Faster search structures for large vocabularies
  • Context-aware correction mechanisms
  • Support for regional languages

These ideas are exploratory and may evolve as BASA grows and receives community feedback.

The current implementation prioritizes simplicity, transparency, and predictable behavior.