📄 Arabic Named Entity Recognition using Optimized Feature Sets

Summary

The paper addresses Arabic named entity recognition. Arabic has no capitalisation, omits short vowels and attaches clitics to words, so many surface forms are ambiguous and the data are sparse. The authors argue against one shared feature set for every NE class. Instead, they build a separate classifier for each class, each with its own tuned features, and then merge their outputs.

They compare two discriminative sequence learners, SVM (via Yamcha) and CRF (via CRF++). The candidate features cover:

• token context windows
• prefix and suffix character n-grams
• gazetteers
• POS tags and base phrase chunks from AMIRA
• a nationality list
• fine-grained morphological attributes from MADA

Each feature is first tested on its own. The Fuzzy Borda Voting Scheme then ranks the features per class, and the top N are added step by step. Experiments use eight standard ACE data sets drawn from several genres. The best result is F=83.5 on ACE 2003 Broadcast News.

The study shows that the most useful features differ strongly between classes. It also shows that Arabic morphology helps NER when it is used as separate, well-chosen attributes rather than as a single block.

Key points

• Features include CXT windows, LEX1–LEX6 (the first and last one, two and three characters of a token), and GAZ lists for persons, locations and organisations. The location list was extended by crawling Arabic Wikipedia from its "Countries of the world" page, followed by stop-word filtering and manual checking.
• AMIRA-1.0 provides POS tags (reported 96.2% accuracy, 25 tags) and BPC (reported F=96.33). MADA supplies morphology across 14 dimensions.
• Feature rankings vary sharply by class. BPC ranks 4th for PER but 13th for GPE and ORG.
• Among the morphological features:
  • Definiteness appears in every class's optimal set.
  • Aspect helps every class except PER.
  • Gender and number raise recall for PER and GPE.
  • The person feature mainly helps ORG and FAC.
• On Weblogs, scores rise from a frequency baseline of 27.66 to 55.32 with a CRF baseline and 57.3 with the best per-class setup. Combining per-class CRFs that use all features drops to 27.36 there, whereas SVM handles the sparse data more robustly.

Interesting ideas

• I1. Each entity class gets its own feature subset, chosen by aggregating per-feature rankings with a fuzzy Borda vote; this beats forcing one compromise feature set on every class.
• I2. Gazetteer matches count only when the whole entity is in the list, but person names are also split into single-word entries, because first and last names recombine freely while place and organisation names do not.
• I3. Morphological attributes such as definiteness, gender and number are fed as separate features instead of one bundled tag, so the learner can pick the ones that matter.
• I4. Fixed-length prefix and suffix character n-grams give a cheap orthographic signal for unseen words.
• I5. The error analysis links failures to class rarity (4 WEA vs. 1707 PER instances) and to inconsistent gold annotation, not only to the model.

For Mova

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

• I1 (learn): Per-label feature or rule selection could be applied when training role or relation classifiers, so each label keeps only the evidence that actually helps it.
• I2 (world): Entity lookup against Wikidata/WordNet lists could use exact multiword matching for places and organisations and partial-name matching for people, which also eases linking later "Clinton"-style mentions.
• I3 (frames): Coreference could score gender, number and definiteness agreement as distinct, explainable signals rather than as one opaque morphology match.
• I4 (tokens): The en tagger could back off to explicit first-k and last-k character features for out-of-vocabulary words, keeping each tagging decision traceable to a named feature.
• I5 (learn): Evaluation reports could list per-class counts and flag inconsistently labelled gold items, so weak scores on rare classes are not misread as parser regressions.