BroadRouter

Taa marbuta

Updated

Taa marbuta (ة, U+0629) is an Arabic letter that appears only at the end of a word, often marking feminine nouns and adjectives. It looks like heh (ه, U+0647) with two dots, so typed text mixes the two, and many search normalizers map ة to ه.

Also called: ta marbuta, teh marbuta, taa marbouta

مدرسة and مدرسه are the same word, “school”, to most Arabic readers. To a string comparison they are different. Taa marbuta is why.

How it works

The Unicode Standard lists TEH MARBUTA (U+0629) as a trailing character, one that occurs only at the end of a word or morpheme. It is right-joining, so it connects only to the letter before it. Its shape is the shape of HEH (U+0647) at the end of a word, plus two dots:

Character Code point Unicode name Final form after ب
ة U+0629 ARABIC LETTER TEH MARBUTA بة
ه U+0647 ARABIC LETTER HEH به

Because the two differ only by two dots at the end of a word, both spellings turn up in typed text. Bouamor et al. (2018) note that while Modern Standard Arabic has a standard orthography, the dialects do not, and people often spell words to reflect their pronunciation. So in informal and dialectal Arabic text, expect a final ه where a dictionary would write ة.

Search normalizers fold the two. CAMeL Tools’ normalize_teh_marbuta_ar is a plain s.replace('ة', 'ه'), and Lucene’s ArabicNormalizer lists “normalization of teh marbuta to heh” as a rule.

Why it matters

Recall. Folding ة into ه lets a query for مدرسه find documents that spell مدرسة, and the reverse. It is one of the standard rules of Arabic text normalization, alongside alef and hamza folding.

Merged words. The rule also collapses words that differ only in this letter:

Original Meaning After ة to ه
كرة ball كره
كره hatred كره

For search, the occasional false match is usually a fair price. For entity matching, spelling correction or legal text, keep the distinction and compare originals.

Scoring. In speech-to-text evaluation, a transcript that writes ه where the reference has ة counts as a word error unless both are normalized the same way. Decide before you score, and state it.

On BroadNet Inference

BroadNet’s API does not normalize text. Chat models receive and return ة and ه exactly as written, and BroadNet Transcribe Arabic returns ordinary Arabic text. Apply the ة rule in your own index or scoring code, on both sides.

If you ask a chat model to correct spelling, check its edits in code: the Arabic spelling correction post shows a diff-based check. For indexes, the Arabic RAG search post and the normalization guide show where the rule fits. You can try the rule on your own strings with the Arabic text normalizer.

Questions

Why not map ه to ة instead?

Because ه is a common letter in its own right, including the pronoun suffix meaning "his". Mapping ة to ه is one-way and safe to apply everywhere; the reverse would invent letters.

Does the ة to ه rule apply in the middle of a word?

Taa marbuta only occurs at the end of a word or morpheme, so in practice the rule only changes final letters. CAMeL Tools replaces every ة without checking position.

Should spelling correction restore ة?

It can, but whether a final ه should be ة depends on the word and sometimes the dialect. Have the model correct it, then check the change in code and in review.

Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity