📄 Using LSTM and GRU With a New Dataset for Named Entity Recognition in the Arabic Language

Summary

This work addresses Arabic named entity recognition, where labeled, structured data is scarce. Existing Arabic NER corpora mostly cover politics, economics, and medicine. Arabic also makes NER harder in several ways: it has no capitalization cue, and optional diacritics (Tashkil) and nunation (Tanween) mean one word can have several different digital spellings.

The authors assembled a new corpus covering seven subject areas: news, sport, and cooking, along with technology, medicine, history, and geography. Every token is labeled in the BIOES scheme over nine entity categories. Before training, the text is normalized by stripping diacritics and nunation.

They then trained two small sequence taggers in Google's Trax library on Colab. Each tagger has a word embedding, one LSTM or GRU layer, a ReLU dense layer, and a LogSoftmax output. Both reach roughly 80% per-token label accuracy, with LSTM slightly ahead of GRU.

The main value of the paper is the multi-domain resource for an under-served language. The modeling itself is a simple baseline.

Key points

• Corpus size and layout: 143 Excel files, 2,104 sentences and 36,380 words. The split is 31,586 words for training, 3,107 for validation and 1,687 for test. Each row has four columns: file_name, sentence, word, tag.

• Label scheme: nine entity types (PER, LOC, GEO/GPE, TIM, PRO, ORG, DIS, MISC) in BIOES format, which gives 37 output classes. For example, a six-word name such as "Muhammad bin Ismail bin Jaafar al-Sadiq" is tagged B-PER, I-PER ×4, E-PER.

• Model setup:

  • 50-dimensional embeddings and 50 recurrent units.
  • Adam optimizer with alpha 0.01 and a cross-entropy loss.
  • About 609k trainable parameters for LSTM and 604k for GRU.

• Accuracy:

  Model	Iterations	Validation	Test
  LSTM	500	81.86%	80.24%
  LSTM	1000	79.88%	80.06%
  GRU	500	78.86%	77.78%
  GRU	1000	76.26%	75.38%

  Doubling the iterations did not improve either model.

• Training time: 88 s and 104 s for LSTM, 90 s and 114 s for GRU (500 and 1000 iterations). The evaluation metric is per-token correctness, not entity-level F1.

Interesting ideas

• I1. BIOES adds explicit End and Single tags, so the boundaries of long multi-word names are recoverable directly from the labels instead of being inferred from where an I-run happens to stop. — For Mova (world): Mova could store entity mentions with BIOES-style boundaries, so multi-token names like "Sir Arthur Conan Doyle" become one world entity with exact spans for coreference to point at.
• I2. Spelling variants that humans read as identical (here, diacritized vs. bare forms) are folded into one canonical key before learning, so they share statistics. — For Mova (tokens): the en crate could keep the surface form but add a normalized lookup key for curly vs. straight quotes, accents, ligatures and British/American spellings, so lemma and lexicon lookups don't split one word into several.
• I3. The corpus is deliberately spread across several unrelated domains, and each row carries a source-file column, so a model can be checked for domain breadth rather than tuned to news text. — For Mova (learn): silver data could be tagged by domain and source (fables, school tests, word problems), with scores reported per domain to catch a parser that only works on one genre.
• I4. Reporting the same model at two training lengths, together with training time, shows that extra training bought nothing and only cost time. — For Mova (learn): training runs could routinely log an early and a late checkpoint, plus cost, so that early stopping is chosen from evidence.
• I5. The headline score is per-token accuracy, which in NER is easy to inflate because most tokens are outside any entity; this is a reminder to pick span-level metrics. — For Mova (learn): Mova's evaluation should score whole entity spans and include an "everything is O" baseline as a negative control, so a tagger can't look good by predicting the majority tag.