DarijaDZ_SA

DarijaDZ_SA is a sentiment-labeled subset of the DarijaDZ corpus of Algerian Darija text. Each document is labeled with one of four sentiment classes: POS (positive), NEG (negative), NEU (neutral), or MIX (mixed/both positive and negative).

---

Dataset Description

Motivation

Sentiment analysis for Algerian Darija is held back by the lack of labeled data. DarijaDZ_SA provides a labeled sample drawn from DarijaDZ to support:

• sentiment classification model training and evaluation
• fine-tuning word/sentence embeddings for sentiment
• Darija NLP research
• studying sentiment expression in informal, code-switched, and multi-script text

---

Dataset Statistics

Metric	Value
Documents	18,275

Label distribution

Label	Count	Share
POS	8,786	48.1%
NEG	4,635	25.4%
NEU	4,539	24.8%
MIX	315	1.7%

---

Labeling Methodology

Documents were sampled from DarijaDZ with a stratified design to ensure adequate representation across script variants (Arabic script, Latin/Arabizi, and mixed/code-switched).

Labels were produced by 5 independent annotators, each labeling a unique portion of the sample, with a shared overlap subset labeled by all 5 for inter-annotator agreement checks. Overlap items were resolved by majority vote (at least 3 of 5 annotators agreeing); items without a majority were resolved through manual review.

Quality checks

• Inter-annotator agreement (Fleiss' kappa, computed on the shared overlap subset): 0.684
• Human evaluation: a stratified sample of agreed-upon labels was independently rated by a human reviewer, blind to the original label. Agreement with the dataset's labels: Cohen's kappa = 0.702 ("substantial agreement").

---

Dataset Format

Distributed as a single JSONL file with the following schema:

Field	Type	Description
id	string	Unique document identifier
text	string	The document text
label	string	One of POS, NEG, NEU, MIX

from datasets import load_dataset

dataset = load_dataset("nasrellahkharroubi/DarijaDZ_SA")

print(dataset)

---

License

This dataset is released under the CC BY 4.0 license.

---

Citation

If you use DarijaDZ_SA in your research, please cite:

Kharroubi Nasrellah.
DarijaDZ_SA: Sentiment-Labeled Algerian Darija Dataset.
2026.