📄 Automatic Correction of Arabic Text: a Cascaded Approach

Summary

The paper describes QCRI's entry in the QALB-2014 shared task on automatically correcting Arabic text, using the QALB dataset. The system splits the work into word-level error correction and punctuation recovery. The first correction component treats spelling repair as transliteration mining. It learns character-sequence mappings from GIZA++-aligned pairs of erroneous and corrected words, proposes in-lexicon candidates, and reranks them with a trigram language model. The second component is a sequence of hand-designed rules for error types the first handles poorly: dialectal words, alef/ya/ta marbouta confusions, wrong splits and merges, letter elongation, and dialectal verb prefixes. Each rule fires only when a bigram LM prefers the result. For punctuation, the authors compared a frequency-threshold heuristic with a CRF over word and POS windows. Running the character model first and then the rules worked best, reaching an official F-measure of about 0.63. Adding punctuation recovery consistently hurt the scores. The work shows that cheap, LM-verified cascades can be competitive for noisy-text normalization.

Key points

• Language resources. The lexicon and LMs come from 234,638 Aljazeera articles covering 10 years plus all of Arabic Wikipedia. Together they hold 576M tokens, of which 1.6M are unique.

• Character-level model.

  • It is built on the El-Kahky et al. (2011) transliteration-mining baseline.
  • Training used roughly one million aligned word pairs, symmetrized with grow-diag-final-and.
  • It generates the top 10 candidates for each word.
  • The best interpolation weight on the LM was λ = 0.6–0.7, meaning the LM matters more than the mapping probability.

• Development-set F-measure for correction.

  Setup	F-measure
  Character-level	0.574
  Case-specific	0.587
  Character → case cascade	0.615
  Case → character cascade	0.603

• Punctuation recovery alone. Both methods scored F = 0.204.

  • Statistical model: precision 0.306, recall 0.153.
  • CRF (CRF++): precision 0.373, recall 0.141.

• Official runs.

  • QCRI-1 (cascade only) was best, with F = 0.6343 and precision 0.717.
  • The runs that added statistical punctuation recovery dropped to 0.6157 and 0.5996.

Interesting ideas

• I1. Spelling correction is reframed as transliteration mining, where learned character-chunk rewrites can only generate words that exist in a large clean lexicon; this keeps every candidate a real word, and the mapping table itself explains why the correction happened.
• I2. Every rule-based edit is gated by a language model, applied only if the sentence scores higher with the change than without it; this turns many aggressive heuristics into safe proposals.
• I3. The cascade order matters: broad statistical correction first, then targeted rules for the leftover error classes, beat the reverse order.
• I4. Candidate word splits are restricted to linguistically plausible boundaries (here, after non-connecting letters) and to words of a minimum length; this sharply reduces false splits.
• I5. A component that looks useful in isolation (punctuation recovery) lowered end-to-end scores, so each module must be tested inside the full pipeline.

For Mova

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

• I1 (tokens): Typo and variant normalization before tagging could use learned character rewrites constrained to the en lexicon, and each fix would be logged with its mapping probability.
• I2 (learn): Self-induced repair rules could be accepted per instance only when an independent scorer, such as witness-parser agreement, improves over the unrepaired version.
• I3 (tokens): Normalization stages should be ordered from general learned models to narrow handcrafted fixes, and the ordering should be checked empirically rather than assumed.
• I4 (tokens): Fixing run-together English tokens (e.g., missing spaces in noisy text) could restrict split points to positions where both halves are lexicon words and the LM favours the split.
• I5 (learn): Every new pre-processing step should be evaluated by its effect on downstream tree and answer accuracy, not only by its standalone metric.