Building a More Useful Hebrew Transliteration Scheme

When the Babel Street Match (formerly Rosette) team set out to build a Hebrew-to-Latin character translator, one of the first considerations was: Which Hebrew transliteration standard should we use? As the joke goes, “Standards are great because there are so many to choose from.”

The existing Hebrew transliteration standards, ISO 259-2:1994 and UNGEGN (United Nations Group of Experts on Geographical Names) — which Match accesses via UNGEGN’s implementation by the ICU open source library — were created for scholars for whom unambiguous scientific transliterations were the top priority. ISO 259-2 was created to handle the processing of bibliographic information. The UNGEGN uses transliteration standards in service of its mission to compile and disseminate global statistical information.

On the other hand, most Match users are looking for Hebrew transliterations that are more compatible with those that show up in a database, or would be used in searches, and are pronounceable. To fulfill these specific requirements, Babel Street created the “folk” transliteration scheme.

The Babel Street scheme is more user-friendly, as it prioritizes pronounceable transliterations and doesn’t contain diacritic marks. It should be noted, however, that this scheme has ambiguity, so it is not possible to roundtrip — i.e., convert from Hebrew to Latin and back to Hebrew, with fidelity. In Hebrew, some letters are pronounced the same, so the Babel Street scheme will map to the same Latin letter in the transliteration. For example, the folk transliterator maps the Hebrew letters “ט” and “ת” to “t,” a many-to-one mapping.

Two difficult Hebrew transliteration cases are worth noting.

1. Difficulties vocalizing Hebrew

First, Hebrew is regularly written without nikud, which are marks that indicate how to pronounce words or may stand for vowels. Native speakers can read Hebrew without nikud because of the word’s context, but Romanized Hebrew is always written with vowels, so the transliteration involves some guessing at the correct vowels. Even if the guessed vowel may occasionally be incorrect, Match should map all the consonants from Hebrew into Latin. There is, however, the case where if the wrong vowel is guessed, an adjacent consonant may also be incorrectly transliterated.

Suppose the correct spelling of a Hebrew word included a diacritic hirik for the sound “i,” followed by the yod, which is transliterated as “y.” In Hebrew, hirik+yod appears as אִי .

Scientific transliteration standards map this combination of hirik+yod (אִי) to “iy” (i.e., following the principle of mapping one character to one character), but the Babel Street transliteration will avoid the “iy” mapping and just use “i.”

Hebrew names with Hirik+Yod

On the other hand, if the correct vowel was patah ( ַ ) with yod, it should be transliterated to Latin “ay,” but Match might make a mistake in some cases and output Latin “iy” instead. In Hebrew, patah+yod appears as אַי .

In other cases, the folk transliterator will just output Latin “i,” which is wrong because it deletes the consonant due to guessing the wrong vowel.

Hebrew names with Patah+Yod

2. The ambiguity of the Shva

The vowel character shva (two stacked dots under a character) has two possible pronunciations: as a vowel or silent. But software cannot distinguish between when the shva is voiced or silent with 100% certainty. When pronounced as a vowel, the shva maps to “e,” but when silent, it should not map to any character. The ICU/UNGEGN always transliterates shva as “ĕ.” The ISO 259-2 standard always deletes it. Babel Street transliterates the shva as “e” or deletes it, depending on the guessed pronunciation.

The Match algorithm makes a best guess about the shva’s pronunciation based on the allowed phonetic clusters in Hebrew. Additional exceptions to the rules are wired in to allow for cases when a borrowed foreign word in Hebrew may violate the phonetic cluster rules.

Hebrew names with silent Shva

Hebrew names with voiced Shva

In brief, the Babel Street folk transliteration scheme attempts to balance fidelity with how words are pronounced in Hebrew, while producing name translations that will resemble what people type into search and database systems.