📄 Arabic Citation Parsing using Part of Speech and Named Entity Recognition

Summary

The paper tackles citation element extraction (CEE) for Arabic: given a reference string, label its parts (author, title, publisher, location, date, pages, translator and so on) and decide whether the text is a citation at all. Citation managers and parsers that handle English well, such as AnyStyle and Zotero, offer almost no Arabic support. Editors therefore restyle and verify Arabic references by hand.

The authors used AnyStyle as a baseline. They then tried a preprocessing-only variant, annotated a new dataset (AraCiteD), and retrained AnyStyle on its original data plus AraCiteD. Their final system, ACEX, combines several signals:

• the retrained model's labels,
• regular expressions,
• Wojood-NER entities,
• CAMeL POS tags,
• a heuristic pass for tokens that remain unlabeled.

ACEX reaches F1 between 0.80 and 0.95 on the important element types. In a deployment at a publishing house it cut citation editing time sharply, and it also handles footnotes that mix several citations with ordinary prose. This matters for Arabic publishing workflows and for building citation networks and impact analysis of Arabic research.

Key points

• AraCiteD: two trained annotators labeled 867 Arabic references with 4,714 elements, taken from 111 articles that follow a Chicago-17 style adapted for Arabic. An expert editor corrected errors in under 6% of elements. The data is released in AnyStyle JSON format.
• Systems compared:
  • AnyStyle, the baseline.
  • AnyStylePre, which normalizes input without retraining.
  • ArAnyStyle, retrained on the original data augmented with AraCiteD.
  • ACEX, the full fusion pipeline.
• External tools: Wojood-NER (21 entity types, above 85% precision and recall) and the CAMeL toolkit for POS tagging.
• Use-case 1 results (database creation), F1:
  • author and title: .99
  • publisher: .95
  • location: .90
  • editor and translator: 1
• Editing time: roughly 2 minutes with ACEX versus about 15 minutes by hand, an 86% reduction. Remaining weak spots are long titles (sometimes truncated) and tangled volume/issue combinations.

Interesting ideas

• I1. Before touching the model, they first normalized the input (punctuation, numerals, letter forms) into the shape an existing model expects, which gives a cheap baseline gain without retraining.
• I2. Instead of training only on the new domain, they kept the original training set and added a small annotated in-domain set, so old coverage is preserved while the new style is learned.
• I3. A statistical sequence labeler is fused with independent signals (regex patterns, NER, POS, cue keywords), each covering the others' blind spots.
• I4. Tokens left unlabeled after the first pass are tracked explicitly as "leftovers" and resolved by a dedicated heuristic, with the plan of feeding them back into the model later.
• I5. Some fields are reliably announced by a trigger word or its abbreviation (for example, page numbers by a "page" keyword), so cue-driven extraction handles them with high precision.

For Mova

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

• I1 (tokens): add a normalization front layer that maps noisy or non-standard punctuation and number forms into the conventions our tagger was trained on, and log every rewrite so it stays explainable.
• I2 (tree): when adapting gparse or the perceptron parser to a new genre, fine-tune on a mix of the base treebank and a small domain treebank rather than on the domain data alone.
• I3 (learn): treat regex, NER and POS cues as extra witnesses next to the witness parsers when building silver data, and accept labels where they agree.
• I4 (frames): collect words not covered by any frame or role as an explicit leftover list. That list could drive a repair pass and become training targets, instead of being silently dropped.
• I5 (world): use explicit cue words (units, "page", "volume", "edition") to type numeric values into measures and quantities before falling back to learned guesses.