📄 Translating English Names to Arabic Using Phonotactic Rules

Summary

The paper addresses inconsistent Arabic spellings of English personal names on official documents. Different spellings of one name create duplicates that can harm security and property-rights records, and existing standards are hard for non-specialists to apply. The authors propose a rule-based transliteration pipeline:

• An English name is first converted to its pronunciation using a name database and the CMU Speech Pronunciation Dictionary.
• The English phonemes are mapped onto Arabic phonetic symbols with full vowel marks (diacritics).
• A set of eight hand-written phonotactic rules then turns this vocalized form into ordinary undiacritized Arabic script. The rules cover word-initial and word-final vowels, the NG cluster and doubled consonants. The system also handles right-to-left output.

For evaluation, 100 names were scored on a 0–10 scale. Partial credit drops with each mismatching symbol. Outputs were compared against Google and Bing, and total agreement scores were about 81–82%. The work matters as an early, fully explainable, linguistically grounded alternative to opaque statistical transliteration of names into Arabic.

Key points

• The pipeline is English name → English phones (CMU dictionary, 39 phonemes) → diacritized Arabic phones → undiacritized Arabic script. An example is OBAMA → OW2-B-AA1-M-AH0.
• The name database holds over 133 thousand names and a phonetic inventory of 84 symbols. Dictionary lookup is the primary route, and a syllable-based model is mentioned as a fallback for unknown names.
• There are eight phonotactic rules. Each is a context pattern with word-start/word-end markers, consonant/vowel classes and optional elements, and it rewrites to one or more phonemes or to nothing.
• Evaluation used 100 test names, each scored out of 10. A difference of one symbol caps the score at 9, two symbols at 8, and so on.
• Comparison with Google gave 41 exact matches and a total score of 815 (81.5%). Comparison with Bing gave 40 matches and 821 (82.1%). These figures measure agreement with those systems rather than accuracy against an independent gold standard.

Interesting ideas

• I1. Pivot through sound, not spelling: transliterate via phonemes from a pronunciation dictionary instead of mapping letters directly, so irregular English orthography stops leaking into the target script.
• I2. Rich internal form, lossy surface last: build a fully specified intermediate (vocalized Arabic) and strip information such as diacritics only at the final rendering step.
• I3. Context rewrite rules with boundary markers and empty outputs: each rule names its left/right context, uses symbol classes (C, V, Cn, optional items) and may delete a segment.
• I4. Lexicon first, rules for out-of-vocabulary items: look up known items in a large dictionary and fall back to general rules only for unseen ones.
• I5. Graded partial credit by symbol differences: instead of binary right/wrong, the score drops by one point per mismatching symbol, which separates near misses from total failures.

For Mova

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

• I1 (back): when rendering English proper names in other languages through the meta-world, a CMU-style phoneme pivot could give consistent and explainable name forms.
• I2 (back): generation should keep fully specified morphological and semantic features in the graph and drop them only during surface realization, which keeps every step inspectable.
• I3 (tokens): the same compact, auditable rule format could drive morphology and lemmatization exceptions in crate en, with every applied rule logged as an explanation.
• I4 (tokens): the lemmatizer and tagger can prefer curated lexicon entries and fall back to rules for OOV words, flagging which path produced each answer.
• I5 (learn): evaluation of string outputs (lemmas, generated names, compressor round-trips) could use similar graded scores so that regressions are visible even when exact-match rates stay flat.