📄 Cross Linguistic Name Matching in English and Arabic

Summary

The paper tackles the task of recognising when a personal name written in Latin letters and a name written in Arabic script refer to the same person. Ordinary string similarity does badly here for two reasons. Both languages follow inconsistent transliteration habits, and Arabic writing usually leaves out short vowels. Gaddafi/Qadhafi is a typical case with many English spellings. The authors start from a normalized Levenshtein edit distance computed after transliterating the Arabic with Artrans. They then add several enhancements:

• character equivalence classes (CEQ), which let one Arabic letter count as a match for any of several English letters;
• rules drawn from an Arabic translator's experience;
• a separate pass for "ch";
• keeping only the vowels that both strings share;
• light stemming of attached Arabic prefixes.

On a NIST TDT name set, CEQ alone raises the F-score from about 48% to about 72%, and all enhancements together reach nearly 80%. The extra cost is only a small constant factor. The broader lesson is that linguistically informed, one-to-many substitution rules can do much more for cross-script matching than generic fuzzy metrics.

Key points

• Baseline data: 106 Arabic and 105 English newswire texts yielded 255 Arabic and 408 English names, of which 29 were shared.
  • Several SecondString metrics were compared: Smith-Waterman, SLIM, Jaro, and Needleman-Wunsch (used as Levenshtein).
  • The Levenshtein variants performed best.
• Normalization: the edit distance is divided by the summed length of both strings and subtracted from 1. Pairs scoring under 0.875 were rejected in the baseline.
• Results on the baseline set: the full set of enhancements improved the F-score by 22%. Implementations were written in perl and Java.
• TDT set:
  • 577 Arabic and 968 English names from roughly 250 news articles, with 272 common names and 641 correct matches.
  • Variants multiply: "Edmond Pope" alone yields 36 correct pairs.
• Thresholds and recall:
  • With CEQ, performance peaks at a threshold near 85%; without CEQ, it peaks near 70%.
  • Recall rises from 57% (baseline) to about 67% with CEQ and to 82% with all enhancements.
  • Complexity grows from |s|·|t| to K(|s|·|t|) with K ≥ 3.

Interesting ideas

• I1. Instead of collapsing letters into coarse classes as Soundex and Editex do, the substitution cost lets one symbol legitimately match a set of alternatives; overlapping sets rarely cause false matches because a real input string is internally consistent (e.g., one dialect's reading of q/j).
• I2. Expensive fallbacks such as prefix stripping are run only when a first score falls in a gray zone between a lower bound and the acceptance threshold, which keeps them from causing over-matching elsewhere.
• I3. The authors sweep the match threshold for each combination of enhancements and plot precision, recall and F; this shows that structured equivalences allow a stricter threshold, not just higher recall.
• I4. Retaining only the vowels attested on both sides aligns a vowel-poor representation with a vowel-rich one before scoring, so the comparison stays fair.

For Mova

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

• I1 (world): Linking entity mentions to Wikidata or knowledge-graph nodes could use explicit, inspectable equivalence sets for name variants, so every match can be explained by the specific rule that licensed it.
• I2 (tokens): Lemma lookup for unknown words could try affix stripping only when direct matching nearly succeeds, logging the fallback as an explicit step.
• I3 (learn): Evaluating new repair rules or silver-data filters as ablation grids across thresholds would reveal whether a rule truly sharpens decisions or merely loosens them.
• I4 (frames): Coreference between name mentions could use a similar asymmetric alignment for abbreviated or partial name forms before comparing them against full mentions.