babelcode 0.1.1

Normalize any language identifier to canonical ISO 639-3 + ISO 15924 form

babelcode

Normalize any language identifier — ISO 639-1 (en), ISO 639-3 (eng), BCP-47 (zh-Hans), NLLB-style (eng_Latn), WikiPron filenames (wp_eng_latn_us), CHILDES corpus names (EnglishNA), or plain English (German) — into a single canonical form:

{iso_639_3}_{iso_15924}

For example: eng_Latn, arb_Arab, cmn_Hans, jpn_Jpan.

Review the mapping decisions used in this project.

Installation

pip install babelcode

Quick start

from babelcode import BabelCode

bc = BabelCode()

bc.normalize("en")          # → "eng_Latn"
bc.normalize("zh-Hans")     # → "cmn_Hans"
bc.normalize("ar")          # → "arb_Arab"  (macrolanguage → preferred individual)
bc.normalize("wp_deu_latn") # → "deu_Latn"  (WikiPron format)
bc.normalize("Farsi")       # → "pes_Arab"  (English name)
bc.normalize("EnglishNA")   # → "eng_Latn"  (CHILDES corpus name)

Singleton shortcut

from babelcode import get_instance

bc = get_instance()  # cached singleton — same instance every call

Accessors

bc.iso639_3("eng_Latn")        # → "eng"
bc.script("eng_Latn")          # → "Latn"
bc.bcp47("eng_Latn")           # → "en"
bc.name("arb")                 # → "Standard Arabic"
bc.scripts("srp")              # → ["Cyrl", "Latn"]
bc.is_macrolanguage("ara")     # → True
bc.macro_members("ara")        # → ["arb", "arz", ...]

Script detection from text

from babelcode import detect_script

detect_script("مرحبا")     # → "Arab"
detect_script("Привет")    # → "Cyrl"
detect_script("こんにちは")  # → "Jpan"

Batch operations

bc.normalize_list(["en", "de", "unknown", "fr"])
# → ["eng_Latn", "deu_Latn", None, "fra_Latn"]

bc.build_mapping(
    source_codes=["en", "de"],
    target_codes=["eng_Latn", "deu_Latn", "fra_Latn"],
)
# → {"en": "eng_Latn", "de": "deu_Latn"}

GlotLID Compatibility Mode

To follow the same decisions made by the GlotLID project use glotlid=True when calling normalize.

from babelcode import BabelCode

bc = BabelCode()

bc.normalize("Farsi", glotlid=True)             # → "fas_Arab"
bc.normalize("ktu_Latn", glotlid=True)          # → "kng_Latn"
bc.normalize("tgl_Latn", glotlid=True)          # → "fil_Latn"

Why iso3_Script instead of BCP-47?

BCP-47 tags (en, zh-Hans) are familiar but they conflate macrolanguages with individual languages, have variable length, and omit the script when it is "obvious" — which is ambiguous for multi-script languages like Serbian.

The {iso_639_3}_{iso_15924} canonical form:

• Always has exactly two components — easy to split and compare.
• Uses ISO 639-3 — one code per individual language, no macrolanguage ambiguity.
• Always includes the ISO 15924 script — no guessing for Serbian (srp_Cyrl vs srp_Latn) or Chinese (cmn_Hans vs cmn_Hant).

Data sources & methodology

babelcode is pre-built from LinguaMeta, a comprehensive open dataset from Google Research that aggregates:

Source	What it provides
ISO 639-3 (SIL)	Three-letter codes for 7 800+ languages
ISO 15924 (Unicode)	Four-letter script codes (Latn, Arab, …)
IETF BCP 47	Standard language tags (en, zh-Hans, …)
LinguaMeta JSON files (~7 500)	Canonical script, macrolanguage membership, English names

Build process

1. Raw data: ~7500 per-language JSON files from LinguaMeta, each containing script associations, name data, and macrolanguage membership.
2. Cache compilation (babelcode-build-cache): Reads all JSON files and produces a single linguameta_cache.json (~550 KB) with pre-resolved BCP-47 → ISO 639-3 mappings, canonical scripts, English names, and macrolanguage membership tables.
3. Runtime: BabelCode loads the cache once and resolves any input format through a cascade of regex matchers and lookup tables.

The cache is distributed inside the package — no network calls at runtime, zero dependencies.

Macrolanguage resolution

By default, macrolanguages resolve to their preferred individual language (e.g. ar → arb Standard Arabic, zh → cmn Mandarin). This can be disabled with resolve_macro=False.

Script inference

When a script is not explicit in the input, babelcode uses this cascade:

1. Script from the input tag itself (e.g. zh-Hans → Hans)
2. text_hint parameter — runs Unicode-range script detection on sample text
3. Canonical script from the LinguaMeta cache
4. Fallback to Latn

Development

git clone https://github.com/omneity-labs/babelcode
cd babelcode
pip install -e ".[dev]"
pytest

Rebuilding the cache

If you update the LinguaMeta data files under src/babelcode/data/_url_nlp_repo/, rebuild the cache:

babelcode-build-cache

License

MIT — Omar Kamali / Omneity Labs