---
license: cc-by-sa-4.0
language:
- ar
pretty_name: BAREC-X
task_categories:
- text-classification
size_categories:
- 1K<n<10K
tags:
- readability
- arabic
- explanations
configs:
- config_name: annotations
  default: true
  data_files: annotations.parquet
- config_name: gpt
  data_files: gpt.parquet
- config_name: gemini
  data_files: gemini.parquet
- config_name: rule_based
  data_files: rule_based.parquet
- config_name: labels
  data_files: labels.parquet
source_datasets:
- CAMeL-Lab/BAREC-Corpus-v1.0
---

# BAREC-X

BAREC-X is the first explanation-annotated Arabic readability dataset. It contains 1,793 sentences from the [Balanced Arabic Readability Evaluation Corpus (BAREC)](https://huggingface.co/datasets/CAMeL-Lab/BAREC-Corpus-v1.0), each with its BAREC readability level (1–19, from early literacy to postgraduate comprehension) and the linguistic reasons that justify that level, 
selected by expert annotators from the BAREC annotation guidelines along six dimensions: spelling, word count, morphology, syntax, vocabulary and content.

The sentences come from the inter-annotator agreement portion of BAREC and were annotated by three native Arabic-speaking educators.
About 300 sentences were labelled by all three; 
for those, the released labels are the reasons at least two annotators agreed on.

The release also includes the labels predicted for the same sentences by GPT-5.5 and Gemini-3-Flash (four-shot prompting) and by a training-free rule-based system, plus the full label inventory.

BAREC-X was introduced in *Benchmarking Fine-Grained Readability Explanations for Arabic* (Rabih, Elmadani, Taha-Thomure and Habash). Project page: https://barec.camel-lab.com/


## Content

| Config | Rows | What it holds |
|---|---|---|
| `annotations` (default) | 1,793 | Human annotations: sentence, readability level, labels. |
| `gpt` | 1,793 | Labels predicted by GPT-5.5 with four-shot prompting. |
| `gemini` | 1,793 | Labels predicted by Gemini-3-Flash with four-shot prompting. |
| `rule_based` | 1,793 | Labels predicted by the training-free rule-based system. |
| `labels` | 148 | The label inventory: code, dimension, level, Arabic description. |


The four sentence tables have the same 1,793 rows in the same order and share the `id` column, so they can be joined on `id` or compared row by row.

All three systems receive the sentence and its gold readability level and return the reasons that justify that level.
The LLM prompts contain a task description, the BAREC guideline excerpt for the sentence's level, and four example sentences with their gold reasons.
The rule-based system extracts its features with CAMeL Tools 1.5.6 and CamelParser 2.0 and returns the activated reasons whose guideline level matches the given level.


## Loading

```python
from datasets import load_dataset

repo = "CAMeL-Lab/BAREC-X"
gold = load_dataset(repo)                 # human annotations
gpt = load_dataset(repo, "gpt")           # also "gemini", "rule_based"
labels = load_dataset(repo, "labels")     # code -> Arabic description
```

## `annotations`

| Column | Type | Description |
|---|---|---|
| `id` | string | Unique 11-digit sentence identifier. |
| `sentence` | string | The Arabic sentence. |
| `readability_level` | int | Readability level, 1–19. |
| `spelling` | list of string | Spelling label codes (prefix `O`). |
| `word_count` | list of string | Word-count label codes (prefix `W`). |
| `morphology` | list of string | Morphology label codes (prefix `M`). |
| `syntax` | list of string | Syntax label codes (prefix `S`). |
| `vocabulary` | list of string | Vocabulary label codes (prefix `V`). |
| `content` | list of string | Content label codes (prefix `C`). |

A dimension with no label is an empty list. A sentence can carry more than one label, in the same dimension or across dimensions.

Example row:

```json
{"id": "10600010006", "sentence": "عالم غريب المعالم", "readability_level": 11, "spelling": [], "word_count": [], "morphology": [], "syntax": ["S11-2"], "vocabulary": [], "content": []}
```

## `gpt`, `gemini`, `rule_based`

These are the outputs of the systems. They contain the same columns as `annotations` without `readability_level`: `id`, `sentence`, and the six label lists holding that system's predictions. The `sentence` text is identical to `annotations`.

## `labels`

| Column | Type | Description |
|---|---|---|
| `code` | string | Label code, e.g. `S10-2`. |
| `dimension` | string | One of `spelling`, `word_count`, `morphology`, `syntax`, `vocabulary`, `content`. |
| `level` | int or null | The BAREC readability level the label belongs to. Null for the "other" codes. |
| `description_ar` | string | The label's description in Arabic. |
| `num_sentences` | int | Number of sentences carrying this label in the human annotations. |

A code has the form `<dimension letter><level>-<index>`: `S10-2` is the second syntax label of level 10. Codes ending in `-O` (`M-O`, `S-O`, `V-O`, `C-O`) mean "other" and have no level.

Sentences per readability level:

| Level | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Sentences | 4 | 12 | 34 | 10 | 49 | 55 | 134 | 130 | 61 | 209 | 89 | 423 | 93 | 337 | 75 | 47 | 23 | 3 | 5 |


## Citation

```bibtex
@inproceedings{rabih-etal-2026-benchmarking,
    title = "Benchmarking Fine-Grained Readability Explanations for Arabic",
    author = "Rabih, Nour  and
      Elmadani, Khalid N. and
      Taha-Thomure, Hanada  and
      Habash, Nizar",
    booktitle = "Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the {A}sia-Pacific Chapter of the {A}ssociation for {C}omputational {L}inguistics",
    month = nov,
    year = "2026",
    address = "Hengqin, China",
    publisher = "Association for Computational Linguistics",
}
```