I just checked out the Scaling to 1000 languages paper from Meta AI, where it mentions uroman as a tool to romanize text across languages. How useful is this in the field of linguistics? What are the use cases?

I ask because, from what I see you lose tones (for Chinese), stress, vowel subtleties, etc., you lose a lot when you do this sort of romanization. So while it may appear readable, and you might get away with a 50% or 70% correct pronunciation using it (a rough ballpark pronunciation it seems), it seems like it is not very useful. But maybe there are other computational or linguistic uses for such a tool, maybe being able to seemingly read it is just a tangential benefit?

Wondering if you could shine a light on that, why you would need such a tool.

  • 2
    While not useful for linguiats, such a tool is useful for lay people (tourists, traders, soldiers ...) Commented Jun 5, 2023 at 8:33
  • 2
    How about IPA??
    – fdb
    Commented Jun 5, 2023 at 9:36
  • 1
    Many languages (e.g, Hmong) use final consonant letters to indicate tones. This works well for those who know the language, and after some training, for those learning the language, too. But Hmong has monosyllabic words with several possible register tones; this wouldn't work for (e.g,) downstep tone languages with polysyllables.
    – jlawler
    Commented Jun 5, 2023 at 13:20
  • None of the Chinese would work, would it? Really? Or Vietnamese or any tone language. It's probably only useful for languages with different alphabets....like Russian.
    – Lambie
    Commented Jun 5, 2023 at 18:37

2 Answers 2


This is not especially useful for linguists, because it sacrifices a lot of linguistically-relevant information. See user6726's answer for more on this.

However, it can be very useful for lay people who are less interested in full accuracy and more interested in having a name they can use to refer to something. Most English-speakers can't pronounce the tones of Mandarin accurately, and wouldn't know what the diacritics in Běijīng meant even if they could. However, "Beijing" is a far more useful name to someone who can't read hanzi than 北京. The former is not entirely accurate, but the latter is nothing but meaningless symbols to them.

And while linguists studying Mandarin can be expected to learn how to read and pronounce it with some level of accuracy, that's not a reasonable thing to require of a traveller trying to find their connecting flight in an airport.


It might be perhaps 1% useful, but perhaps -50% useful.

I still can't read Tigrinya fidäl and definitely cannot write it, but I wrote a conversion program that handles unicode text and converts to conventional Latin letters (bidirectional, so Latin-to-Ethiopic is much more useful to me). The main problem is that there is almost no unicode text in the script, and no realistic prospects for creating any such text corpus (a lack of reliable OCR and a dearth of usable printed text to be analyzed). OTOH there is a lot of material in Telugu which is unicoded, and I could easily write a conversion program if I wanted (I don't do Telugu, so that ends that discussion). My grasp of Devanagari (for Sanskrit) is good enough that I can make reasonable guesses about Hindi, except that there are a lot of details in the orthography-to-pronunciation realm that I don't understand and AFAIK is poorly understood (esp vowel nasalization and schwa deletion). But I also don't do Hindi.

If the goal is to get phonological data on the language, the prospects are extremely poor that any transliteration scheme will provide information that actually cannot be discerned from the writing system. That is even true in a Latin-based writing system like that for North Saami, where you would not know from spelling that <lieđbmi> is pronounced [lĭeðeʔmi], unless you work out a system of orthography-to-pronunciation rules. You can never distinguish the two pronunciations and meaning of orthographic <akini> in Logoori as [akɪ́nɪ́] 'that he plow' and [akɪni] "he plowed", without already knowing the language.

IMO this tool is as useful as a loaded pistol with a hair trigger – it is potentially dangerous. Making up data in Tigrinya by relying on a letter-conversion tool is bad for the field. You might think that ሕቈ is pronounced [ħɨk'ʷʌ] based on transliteration schemes, but it is pronounced [ħɪʁ̆ɔ]. For syntax this might be an acceptable approach when you don't care about actual pronunciation. Specialists have no problem with [ħɨk'ʷʌ] or other such transliteration schemes, but a lot of linguistic analysis nowadays is based on low-exposure computational scraping methods. Facile auto-transliteration is counterproductive unless preceded by standardization of the translation and a clear understanding by the user that the output is not about pronunciation. The sad history of Modern Mongolian linguistic studies in the area of phonology underscores how easily linguists can be misled by confusing transliteration and transcription.

I would say that if a syntactician harvests a line of text written in an obscure script that they cannot read, they don't understand the language well enough to make valid use of the datum.

  • "The sad history of Modern Mongolian linguistic studies in the area of phonology underscores how easily linguists can be misled by confusing transliteration and transcription." What is this segment referring to? Do you have any sources where I can read about this sad history you mention?
    – maritsm
    Commented Jun 10, 2023 at 8:44
  • Svantesson et al have a book which describes the language, based on actual pronunciation, which is somewhat revolutionary, and they also briefly note some prior works. To get the full picture, you would want to read previous grammatical descriptions and derivative words, and see how those works diverge from speech.
    – user6726
    Commented Jun 10, 2023 at 13:43

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.