📄 Converting Romanized Persian to the Arabic Writing Systems

Summary

The paper addresses the conversion of Persian written in Latin letters, which is common in blogs, email and SMS and among diaspora speakers, back into the traditional Perso-Arabic script. The source is Dabire, a phonemic romanization in which each phoneme maps to exactly one letter. The authors build a purely rule-based converter in Xerox XFST. It uses no dictionary and no morphological analysis. The input word passes through three composed finite-state transducers:

1. Syllabification groups the letters into syllables, maximizing onsets.
2. Context rules choose the right letter shapes for vowels, which depend on syllable type and position in the word.
3. Transliteration maps the result to Unicode characters.

The authors report near-perfect output for non-Arabic foreign names. Accuracy drops for Arabic loanwords, compounds and irregular spellings, because the romanization has already lost information the Arabic script needs. The work matters because the system can handle new words and loanwords absent from any lexicon. The authors also present it as the first automatic converter from romanized to Arabic-script Persian.

Key points

• The pipeline is the XFST composition Sfy .o. CR .o. D2P. Syllables are defined as V, VC, VCC, CV, CVC and CVCC, with forbidden consonant clusters (e.g. bp, kq) excluded. The @> replace operator picks the shortest matches from left to right.
• Persian has 24 consonants and 6 vowels: short a, e, o and long â, i, u. Dabire maps these phonemes one-to-one to its letters.
• The authors define P-Script, a subset of the Perso-Arabic script without the Arabic-only letters, which are redundant for Persian. Short-vowel diacritics are available in the intermediate form but are dropped by default.
• The output is close to 100% correct for non-Arabic foreign names and simpler texts such as the Shahname. The authors give four sources of error:
  • letters specific to Arabic,
  • ad hoc compound-writing conventions,
  • multi-role letters such as He (consonant /h/ vs. word-final vowel),
  • irregular spellings, such as a silent Vâv.
• According to the authors, most remaining errors concern style rather than spelling. They are adding lexical and morphological information.

Interesting ideas

• I1. Placing a syllable layer between letters and output gives context rules a structural position (syllable type, word-initial or word-final) to condition on, instead of raw character windows.
• I2. Building the converter as a chain of separately inspectable transducers, each with a documented intermediate form, makes every step traceable, as the stepwise table of examples shows.
• I3. The authors explain errors as information lost upstream rather than as rule failures, separating fixable defects from inherently unrecoverable cases.
• I4. A lexicon-free rule core serves as the fallback for unseen words, while lexical knowledge is planned as an overlay for exceptions.
• I5. Breaking the evaluation down by text type (epic poetry, foreign names, modern prose with loanwords and compounds) reveals where the method degrades.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (tokens): the morphology and lemma rules in crate en could key off syllable or segment structure for spelling changes (doubling, -y→-ie), keeping each decision explainable.
• I2 (back): the generation and compressor path could expose named intermediate layers, so a wrong surface form can be traced to the exact rule that produced it.
• I3 (learn): evaluation could tag errors as "information absent in the input" versus "model mistake," so repair rules are not induced for cases no rule can fix.
• I4 (world): the compiled knowledge (WordNet, Wikidata) can act as an exception layer over general rules, with unknown entities still handled by the rule core instead of failing.
• I5 (answer): reporting scores separately for fables, word problems and school tests would show which genre-specific phenomena break the readers.