[Paper] Algerian Dialect

Source: arXiv - 2512.19543v1

Overview

A new, publicly‑available resource called Algerian Dialect brings 45 k YouTube comments written in Algerian Arabic (Darija) together with fine‑grained sentiment labels. Because dialectal Arabic has been notoriously under‑represented in NLP benchmarks, this dataset fills a critical gap for anyone building sentiment‑aware applications that need to understand North‑African online chatter.

Key Contributions

• Large‑scale, sentiment‑annotated corpus – 45 000 YouTube comments collected from 30+ Algerian news and media channels.
• Five‑point sentiment scale – each comment is manually tagged as very negative, negative, neutral, positive, or very positive, enabling more nuanced modeling than binary polarity.
• Rich metadata – timestamps, like counts, video URLs, and annotation dates are included, opening doors for temporal or popularity‑aware analyses.
• Open licensing – released under CC BY 4.0 on Mendeley Data, allowing unrestricted academic and commercial reuse.
• Baseline experiments – the authors provide initial benchmark results using classic machine‑learning classifiers and modern transformer models, establishing a performance reference for future work.

Methodology

1. Data collection – The YouTube Data API was used to scrape comments from a curated list of Algerian press and media channels. Only publicly visible comments were retained, and duplicate or spammy entries were filtered out.
2. Pre‑processing – Comments were normalized for Arabic script variations (e.g., handling Arabic‑Latin code‑switching common in Darija) and stripped of URLs, emojis, and other non‑textual artifacts while preserving sentiment‑bearing tokens.
3. Annotation – Trained native speakers manually assigned each comment to one of the five sentiment categories. Inter‑annotator agreement was measured (Cohen’s κ ≈ 0.78), indicating reliable labeling.
4. Benchmarking – The authors split the data into train/validation/test sets (80/10/10) and evaluated several models:
   • Logistic Regression & SVM with TF‑IDF features
   • Bi‑LSTM with word embeddings trained on Arabic corpora
   • Pre‑trained multilingual BERT (mBERT) fine‑tuned on the dataset

Performance was reported using macro‑averaged F1‑score to reflect the balanced treatment of all five classes.

Results & Findings

• Baseline performance – Traditional TF‑IDF + Linear models achieved macro‑F1 scores around 0.55, while the Bi‑LSTM reached ~0.62. The fine‑tuned mBERT model delivered the best results, with a macro‑F1 of roughly 0.71, confirming that transformer‑based approaches can effectively capture the nuances of Algerian Darija despite limited pre‑training data.
• Class distribution – The dataset is relatively balanced across the five sentiment buckets, which helps avoid bias toward the majority class (often “neutral” in other Arabic sentiment corpora).
• Metadata utility – Preliminary analyses showed a correlation between comment like counts and positive sentiment, suggesting that metadata can be leveraged for semi‑supervised or weakly‑supervised learning.

These findings demonstrate that Algerian Dialect is both a challenging benchmark (due to dialectal spelling variations) and a fertile ground for testing cross‑lingual and domain‑adaptation techniques.

Practical Implications

• Social‑media monitoring tools – Companies targeting the North‑African market can integrate models trained on this dataset to automatically gauge public opinion on products, campaigns, or political events.
• Customer‑support chatbots – Adding Algerian Darija sentiment detection enables more empathetic, context‑aware responses in conversational agents serving Algerian users.
• Content moderation – Platforms can flag potentially harmful or abusive comments with higher precision by using dialect‑specific sentiment cues rather than relying on MSA‑only models.
• Research acceleration – The open license encourages startups and academic labs to experiment with transfer learning, data augmentation, or multimodal (text + audio) sentiment analysis without the overhead of building a dataset from scratch.

Because the dataset includes timestamps and engagement metrics, developers can also build trend‑aware dashboards that track sentiment shifts over time—useful for PR teams, political analysts, or market researchers.

Limitations & Future Work

• Platform bias – All comments come from YouTube; sentiment patterns on other platforms (Twitter, Facebook, TikTok) may differ.
• Dialectal variability – Algerian Arabic varies regionally and often mixes Arabic script with Latin characters and numerals (e.g., “3” for “ع”). While the dataset captures this mix, models may still struggle with less‑common orthographic conventions.
• Annotation granularity – Five sentiment levels provide nuance, but they do not capture specific emotions (e.g., anger, joy) that could be valuable for affective computing.

The authors suggest extending the corpus to additional social‑media sources, enriching it with emotion labels, and exploring multimodal signals (audio/video) to further boost real‑world applicability.

Authors

• Zakaria Benmounah
• Abdennour Boulesnane

Paper Information

• arXiv ID: 2512.19543v1
• Categories: cs.CL
• Published: December 22, 2025
• PDF: Download PDF