📄 Person Name Entity Recognition for Arabic

Summary

The paper addresses recognizing person names in Arabic text, where NER resources were scarce in 2007. The authors built PERA, a rule-based recognizer with two parts. The first is a set of gazetteers, or name lists, collected from the ACE and Arabic Treebank corpora, government and company name databases, and web lists that were transliterated into Arabic script with Sakhr's Tarjim. The second is a grammar of regular expressions that encodes Arabic naming structure (ism, kunya, nasab, laqab, nisba) and contextual cues such as titles and job words. Recognition runs in two passes: a gazetteer lookup first, then the grammar rules. A "rejecter" filter marks where a name stops and ordinary text begins. Rules and entries also cover common spelling variants, such as hamza, taa marbouta and aleph maqsura alternations. The system was deployed inside a FAST ESP pipeline and evaluated with the hurricane tool on a semi-automatically tagged reference corpus. It exceeded its preset targets. The work shows that careful linguistic engineering can yield usable NER for a low-resource language.

Key points

• PERA combines gazetteers with regular-expression grammar rules and applies them in two cycles: lexicon lookup first, then rules.
• Its auxiliary dictionaries include:
  • job titles (19,245 entries)
  • honorifics (173)
  • country names with spelling variants (923)
  • nicknames and laqabs (8,169)
  • person titles (20)
  • person-indicator phrases (421)
• The reference corpus is about 4MB of text from ACE and the Treebank. It was tagged by a Python script, manually corrected, and split into 46 test sets.
• Average results were 85.5% precision, 89% recall and 87.5% F-measure, against targets of 80% precision and 70% recall. Per-set F-measure ranged from about 84 to 95.
• Some gazetteer entries came from the same corpus used for evaluation. The authors argue the results still hold because the grammar found names that were absent from the lists.

Interesting ideas

• I1. The gazetteer pass runs before the grammar pass, so cheap high-confidence lookups settle easy cases and rules handle the unseen names. This layered design keeps every decision traceable to a list entry or a named rule.
• I2. The grammar models the internal structure of a culture's naming system, with ordered slots such as kunya + ism + nasab chains + nisba, instead of treating a name as an opaque string. This helps both to find name boundaries and to split names into parts.
• I3. Context words that signal a person, such as titles, job words and honorifics, are mined from the phrases around annotated names and stored as their own dictionary. Such cues let the system recognize names it has never seen.
• I4. A dedicated "rejecter" filter decides where a name ends, which separates boundary detection from name detection. Errors become easier to diagnose because each component fails in its own way.
• I5. The authors admit that missing tags in their semi-automatic gold data turned correct outputs into false positives. Gold-annotation gaps can distort evaluation as much as model errors do.

For Mova

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

• I1 (tokens): Mova's tagger could add a lexicon-first proper-name pass before rule-based fallback, so each NNP decision records which source produced it.
• I2 (world): Entity records could store structured name parts (given name, patronymic, epithet, origin), which would support coreference between short and full mentions of the same person.
• I3 (frames): Coreference and role assignment could use mined title/occupation cues to type an unknown mention as a person before choosing pronoun antecedents.
• I4 (tree): Multiword flat/compound name spans in UD parses could get an explicit boundary-check rule that flags or repairs over-extended name chains.
• I5 (learn): Evaluation reports could separately audit a sample of "false positives" against the gold, to estimate how much of the measured error comes from incomplete silver or gold labels.