📄 Recognition and Translation of Arabic Named Entities with NooJ Using a New Representation Model

Summary

The paper presents a rule-based, finite-state pipeline built in NooJ that recognizes Arabic named entities and translates them into French. Its main contribution is a feature-structure representation of NEs that does not depend on domain or lexical category. The design draws on HPSG and on Bourigault's head/expansion idea. Each NE carries a type, an optional trigger word, and a head that can itself be a complete NE, which handles nested entities of unknown length. It may also have an optional ending part introduced by a linking item such as a preposition. Two well-formedness principles, saturation and non-saturation, separate true NEs from fragments that only contain one. Recognition relies on dictionaries and transducers. Translation works in two steps: each feature value is mapped through bilingual dictionaries, and readjustment rules then fix word order and agreement in French. On sports newspaper text the recognizer reaches 94% F-measure, and 97% of translated NEs come out well formed. The approach keeps recognition and translation as separate modules, so each can be reused across domains and language pairs.

Key points

• The representation uses the features Type_EN, Tête_EN (head), Mot_déclencheur (trigger word), Fin_EN (end) and Element_EN (linking item). Tête_EN and Fin_EN may be atomic or recursively structured.
• The sport corpus has 4000 newspaper texts (94.5 Mo) from Assabah, al Anwar, al chourou9 and al ahram, plus Wikipedia. It holds 180,000 NEs, of which 40,000 are place names.
• Sport-domain recognition scores 98% precision, 90% recall and 94% F-measure. When several transducer paths match, the longest one is kept.
• On a second domain, 300 education texts with university institution names, precision stays at 98% but recall falls to 70% (82% F-measure). The authors blame gaps in the domain dictionaries and missing transducer paths.
• Arabic→French translation yields 97% well-translated NEs. The remaining errors come mostly from ambiguous toponyms (e.g., Tunis vs Tunisia), and first names are left for later transliteration.

Interesting ideas

• I1. An entity is modeled as a typed head plus an optional trigger word and an optional removable "ending" (e.g., a locative phrase); the head can recurse into another entity, and an entity only counts as complete when it has a typed, non-empty head.
• I2. Translation first maps each slot of the structure through a bilingual lexicon, accepting an ill-formed intermediate result, and then applies target-language readjustment rules for order and agreement.
• I3. Recognition and translation are deliberately decoupled (transfer instead of a combined "semi-direct" pass), so the recognizer serves any target language and the French side serves any source language.
• I4. When a dictionary offers several translations, a co-occurring adjective selects the compatible one (a "sportive" modifier favors city over country for madina).
• I5. The authors test their resources on a second domain and track silence (missed items) separately from noise (false hits). This exposes coverage gaps that a single-domain score would hide.

For Mova

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

• I1 (world): Entity nodes could be built as nested structures (stadium → king → first name), with "in Riyadh" stored as a detachable location link, and an explicit completeness check before a mention becomes a world entity.
• I2 (back): Translation through the meta-world could translate graph node labels slot by slot, then hand the result to a target-specific realizer, which keeps lexical transfer and surface grammar explainable as separate steps.
• I3 (back): Keeping the analysis graph independent of the output language means generation modules can be added per language without touching parsing.
• I4 (world): Modifier–head compatibility checked against the ontology could pick word senses or entity types in an explainable way.
• I5 (learn): Evaluations could include an out-of-domain split and report missed versus spurious items separately, as a cheap check on how complete the lexicon is.