📄 Arabic Named Entity Recognition: Using Features Extracted from Noisy Data

Summary

Arabic NER is hard because of rich morphology and small hand-labelled corpora, so features that need many examples, like syntactic context, are hard to learn from gold data alone. The authors start from their strong SVM/CRF system (BASE), which uses one classifier per class. BASE relies on shallow cues: a ±1 word window, character n-grams, gazetteers, AMIRA POS and chunk tags, MADA morphology, and a capitalization signal taken from English glosses. EXTENDED adds features drawn from Bikel-parser phrase-structure trees: the headwords and non-terminal labels above each entity. It also adds class-split gazetteers and wider n-gram context lists. To get enough examples of these features, they tag the English side of a hand-aligned Arabic–English parallel corpus with an English NER system and project the labels onto the Arabic side. The noisy result supplies feature statistics only; it is never mixed into the gold training data. On ACE 2003/2004/2005, the gain reaches 1.64 F (ACE 2004 BN), with 1.16 F on weblogs. This shows that noisy cross-lingual data can enrich features without polluting the training labels.

Key points

• The projection corpus is mostly LDC hand-aligned data across five genres, with 941,282 Arabic tokens. It yields 57,290 projected NEs: GPE-like LOC 27,651, PER 17,964, ORG 10,572, FAC 998, VEH 85, WEA 20.
• The English tagger used for projection reports about 92% performance (Zitouni and Florian, 2009). Parsing uses the Bikel implementation of the Collins parser; Arabic parsing accuracy is only in the low 80s.
• New features:
  • Syntactic head-word (SHW): lexicalized parent heads.
  • Syntactic Environment (SE): parent non-terminals.
  • Class Based Gazetteers (CBG).
  • N-gram context (NGC): windows of 1–3 words.
• Using only the first parent non-terminal (SE) as a single feature gave up to about 4 points over the baseline. Adding the second and third parents lowered results by 0.5 and 1.5 F.
• Evaluation covers ACE 2003 (BN, NW), ACE 2004 (BN, NW, ATB) and ACE 2005 (BN, NW, WL), compared against a most-frequent-tag baseline. The best gain is 1.64 F on ACE 2004 BN.

Interesting ideas

• I1. Noisy data as a feature source, not as labels. Use noisy projected data only to gather feature statistics and lexicons, never as training labels, so errors cannot directly corrupt the decision boundary.
• I2. Capitalization borrowed from glosses. Arabic has no letter case, so the system checks whether the English gloss of the chosen morphological analysis is capitalized. A cue missing in one language is taken from a lexicon's link to another language.
• I3. Shallow tree context beats deep context. The single immediate parent label helped most; climbing further up the tree added noise. Context depth should be tuned rather than maximized.
• I4. Tree context for messy genres. On weblogs, local word windows were unreliable, while tree-based context still helped. This suggests syntax-derived features are more robust when word order is chaotic.
• I5. Per-class classifiers with a precision tie-breaker. One classifier per class, with conflicts resolved in favor of the most precise classifier, gives a simple and auditable decision rule.

For Mova

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

• I1 (learn): Silver parses on which the witness parsers disagree could still feed lexical and context statistics, while only agreed trees act as training targets.
• I2 (world): Entity typing could check whether a token's WordNet or Wikidata link points to a proper-name entry, as a second explainable signal beyond surface case.
• I3 (tree): Features for UD relation or NE decisions could start with the immediate head and its deprel, and add grandparents only if held-out tests show a gain.
• I4 (frames): For informal texts such as dialogues in tales, role and entity cues could rely on dependency heads rather than adjacent tokens.
• I5 (world): When several entity-type detectors fire on one span, choosing the historically most precise one keeps the decision traceable and easy to explain.