📄 A New Error Annotation for Dyslexic texts in Arabic

Summary

Arabic has had no standard scheme for classifying the spelling errors of dyslexic writers. Earlier work mostly listed typical mistakes without organising them into a taxonomy. Alamri and Teahan build such a taxonomy, DECA (Dyslexic error classification for Arabic texts). They drew on prior error-analysis studies and on real texts by dyslexic children. DECA groups errors into nine categories (Hamza, Almadd, Confusion, Diacritics, Form, and others) with 37 error types, plus "Other" fallbacks.

The scheme was refined in two stages:

• Pilot annotation: a 5,000-word sample was annotated, which surfaced two missing types.
• Expert check: two special-education teachers reviewed the revised scheme.

The authors then fully annotated BDAC, a 27,136-word corpus of texts by diagnosed dyslexic Saudi students, containing about 8,000 errors. They used a custom Java tool that stores annotations as XML. In the analysis, the "Common errors" category is the largest, and Alif Madd is the single most frequent error type. The work gives Arabic its first dyslexia error resource, useful for pedagogy, research and future spell-checking aids.

Key points

• DECA Version 2 has 9 categories and 37 error types. Version 1 had 35 types; "Repeated letters" and letter-shape-by-position (beginning/middle/end) were added after the pilot.
• BDAC contains 27,136 words and roughly 8,000 errors. The writers are boys and girls aged 8–12 (a few aged 13), and the sources are Saudi primary schools, online forms and parents.
• Error counts by category:
  • Common errors: 2,717
  • Hamza: 1,621
  • Confusion: 1,553
  • Differences and Form: the least frequent categories.
• Alif Madd errors make up 13.43% of all errors. The most frequent single erroneous word is a misspelling of "on", with 64 occurrences.
• Tags use two letters, category then type (e.g. HA), following Alfaifi et al. (2013). A word with several errors gets joined tags such as HA_MA. Each XML record stores the token number, error form, correct form and tag.

Interesting ideas

• I1. Before committing to full annotation, the authors annotated a small sample with the draft scheme and turned the leftover cases into new types; this exposes coverage gaps cheaply.
• I2. Tags are built compositionally (category letter plus type letter), with an explicit "Other" slot inside every category, so unknown cases stay countable instead of being forced into a wrong label.
• I3. A token carrying several independent errors gets several joined tags rather than one blended label, which keeps the statistics per error type clean.
• I4. Every record pairs the erroneous form with its corrected form and token position, so the corpus doubles as training data for a corrector.
• I5. Annotators were domain teachers, and they were asked about both tag correctness and how easy it was to choose a tag; ease of choosing is a useful signal of ambiguous categories.

For Mova

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

• I1 (learn): Run a pilot of any new silver-label or repair-rule category set on a few hundred sentences, logging cases that fit no label, before scaling.
• I2 (tokens): Typo and normalisation categories in the en crate could use hierarchical codes with per-group "other" buckets, keeping unexplained cases visible in evaluation.
• I3 (tokens): When the tokenizer repairs a misspelled word, it could record each edit as a separate reason, making the correction fully explainable.
• I4 (learn): Noisy-input test sets for the parser could store original/corrected pairs with positions, so robustness can be measured against a clean-text control.
• I5 (frames): When LLM teachers or humans assign semantic roles or speech acts, also collecting a difficulty rating could flag frame labels that need clearer definitions.