PhonEx — French Phonics Exercise Corpus for Dyslexia

Dataset Description

PhonEx is a structured corpus of French-language phonics exercises for children aged 6–9 with reading difficulties including dyslexia, annotated for phonological difficulty level and primary error type. It is designed for constraint-aware exercise retrieval — matching a parent's or teacher's free-text description of a child's reading difficulty to the most appropriate remediation exercise.

The corpus supports the CASS (Context-Aware Semantic Similarity) framework — a lightweight safety-aware retrieval system that prevents pedagogically dangerous mismatches (e.g. assigning a difficulty-level-5 paragraph exercise to a child at difficulty level 1).

• Paper: CASS: A Context-Aware Semantic Similarity Framework for Safe Retrieval in Child-Facing Educational Applications (Sadouk and Gadi)
• GitHub: https://github.com/lsadouk/CASS
• Application: Lecture Magique — French phonics learning app for dyslexic children
• Target curriculum: Moroccan primary school CP–CE1 (ages 6–9)

---

Dataset Summary

Property	Value
Total entries	100
Error types	5
Difficulty levels	1–5
Balance	4 entries per error type × difficulty combination (perfectly balanced)
Age groups	CP (6–7), CE1 (7–8), CE2 (8–9)
Language	French
License	CC BY-NC 4.0

---

Supported Tasks

• Exercise retrieval — match a parent/teacher's free-text description of reading difficulty to the most appropriate phonics exercise
• Difficulty classification — predict phonological difficulty level from exercise description
• Error type classification — predict error type from exercise description or parent query

---

Dataset Structure

Fields

Field	Type	Description
id	string	Unique identifier e.g. ex_vc_003
ERR_TYPE	string	Primary error type targeted (see taxonomy below)
DIFF	int	Phonological difficulty 1–5
AGE_GROUP	int	Target age group: 1=CP/6-7, 2=CE1/7-8, 3=CE2/8-9
description	string	2–4 sentence English description of the exercise
child_instruction	string	Simplified French instruction shown to the child
ERR_SECONDARY	string or null	Secondary error type if applicable
language	string	fr
target_population	string	children_with_dyslexia_age_6_9
annotator_agreed	bool	Annotation confirmed by domain reviewer
version	string	Corpus version

Error Type Taxonomy

ERR_TYPE	Description	Count
visual_confusion	Confusion between visually similar letters (b/d, p/q, m/n)	20
vowel_substitution	Substitution of vowel sounds (o→ou, eu→ou, an→en)	20
syllable_omission	Omission or addition of syllables in words	20
letter_reversal	Reversal or transposition of letter order within words	20
blending_difficulty	Difficulty blending phonemes into syllables or words	20

Difficulty Distribution

DIFF	Description	Count
1	Isolated letters or simple CV syllables	20
2	Simple CVC words and two-syllable words	20
3	Two-syllable words in sentences, short texts	20
4	Sentences, reading aloud, connected text	20
5	Paragraphs, fluency, academic vocabulary	20

---

Example Entries

{
  "id": "ex_vc_001",
  "ERR_TYPE": "visual_confusion",
  "DIFF": 1,
  "AGE_GROUP": 1,
  "description": "The child confuses the letters b and d when reading
    simple CVC words like 'bal' and 'dal'. The exercise shows large
    isolated letters b and d side by side and asks the child to point
    to the correct one when the teacher says the sound. Visual cues
    and colour coding help distinguish the two letters.",
  "child_instruction": "Regarde les lettres. Montre-moi la lettre 'b'.",
  "ERR_SECONDARY": null,
  "language": "fr",
  "target_population": "children_with_dyslexia_age_6_9",
  "annotator_agreed": true,
  "version": "1.0"
}

{
  "id": "ex_bd_009",
  "ERR_TYPE": "blending_difficulty",
  "DIFF": 5,
  "AGE_GROUP": 3,
  "description": "Severe blending difficulty at sentence level — the
    child loses sentence meaning while decoding individual words,
    reading in a slow monotone that prevents comprehension. Exercise
    uses prosodic reading support — stress marks, rhythm bars, and
    pause symbols are printed above the text.",
  "child_instruction": "Suis les marques au-dessus du texte: accent fort,
    barre de rythme, pause. Lis avec la mélodie.",
  "ERR_SECONDARY": null,
  "language": "fr",
  "target_population": "children_with_dyslexia_age_6_9",
  "annotator_agreed": true,
  "version": "1.0"
}

---

Usage

from datasets import load_dataset

ds = load_dataset("lamyaa/PhonEx")

# All entries
print(f"Total entries: {len(ds['train'])}")

# Filter by error type
visual = [e for e in ds['train'] if e['ERR_TYPE'] == 'visual_confusion']
blending = [e for e in ds['train'] if e['ERR_TYPE'] == 'blending_difficulty']

# Filter by difficulty level
beginners = [e for e in ds['train'] if e['DIFF'] == 1]
advanced  = [e for e in ds['train'] if e['DIFF'] == 5]

# Filter by age group
cp_level = [e for e in ds['train'] if e['AGE_GROUP'] == 1]

Use with CASS retrieval

# See full code in: https://github.com/lsadouk/CASS

# English description
retrieve_phonex("My child confuses b and d when reading simple words")
# → ex_vc_003 (visual_confusion, DIFF=2)

# French description
retrieve_phonex("Elle saute des syllabes dans les mots longs")
# → ex_so_006 (syllable_omission, DIFF=3)

# Moroccan Darija
retrieve_phonex("Weldi ma iqrach mezyan, kayqra harf harf")
# → ex_bd_001 (blending_difficulty, DIFF=1)

---

Safety Properties

PhonEx is designed with pedagogical safety in mind. The DIFF and ERR_TYPE annotations enable the CASS framework to prevent dangerous mismatches:

• A child at DIFF=1 (isolated letters) will never receive a DIFF=5 exercise (paragraph fluency) — which would cause immediate failure and frustration
• An exercise targeting visual_confusion will never be retrieved for a child with syllable_omission — providing zero therapeutic benefit
• Safety filtering is automatic — no teacher or parent configuration required

Across 6 diagnostic queries:

• C_DIFF (difficulty): 13 dangerous entries suppressed, mean rank drop 25.3 positions, max 69 positions
• C_ERR (error type): 15 dangerous entries suppressed, mean rank drop 55.7 positions, max 78 positions

---

Annotation Guidelines

DIFF scale:

• DIFF=1: Isolated letter or phoneme recognition, single sound discrimination
• DIFF=2: Simple CVC words, two-syllable words in isolation
• DIFF=3: Words in sentences, short connected text, specific error patterns
• DIFF=4: Reading aloud in connected text, under time or social pressure
• DIFF=5: Extended paragraphs, fluency, academic vocabulary, multiple simultaneous errors

ERR_TYPE selection: Each exercise targets one primary error type. Exercises with meaningful secondary error involvement include an ERR_SECONDARY field. The five error types are mutually exclusive as primary categories but share some underlying phonological mechanisms — used in the CASS C_ERR partial overlap scoring (score 0.5 for related types, 0.0 for unrelated).

Partial overlap pairs:

• visual_confusion ↔ letter_reversal (both involve letter-level processing)
• blending_difficulty ↔ syllable_omission (both involve syllable-level processing)
• blending_difficulty ↔ vowel_substitution (both involve phoneme-grapheme mapping)

---

Citation

@article{sadouk2025cass,
  title     = {CASS: A Context-Aware Semantic Similarity Framework
               for Safe Retrieval in Child-Facing Educational Applications},
  author    = {Sadouk, Lamyaa},
  journal   = {Education and Information Technologies},
  year      = {2025},
  publisher = {Springer}
}

---

Dataset Card Authors

Lamyaa Sadouk Ecole Marocaine des Sciences de l'Ingénieur, Casablanca, Morocco

Taoufiq Gadi Laboratoire de Recherche en Mathématique, Informatique et Sciences de l'Ingénieur (MISI), Morocco

License

CC BY-NC 4.0 — Attribution-NonCommercial 4.0 International

You may use and share this dataset for research and educational purposes with attribution. Commercial use is not permitted without explicit permission from the author.