📄 Algerian Dialect

Summary

Algerian Arabic is widely spoken but has few public NLP resources. It differs sharply from Modern Standard Arabic, borrows heavily from French and Berber, and often switches languages and scripts within one comment. Most existing Arabic sentiment corpora cover MSA or use only coarse polarity labels. This data paper describes Algerian Dialect, a corpus of 45,000 YouTube comments. The comments were gathered through the YouTube Data API from over 30 channels run by Algerian news outlets and media organizations, so most of them react to political, social and economic news. Native speakers labeled each comment on a five-point ordinal scale from very negative to very positive. Unclear cases were discussed as a group until the annotators reached consensus. The text is stored without aggressive normalization, and each record includes metadata such as like count, source video and annotation time. Most labels are mild positive or negative, and the two extreme classes together make up under 9%. The authors earlier used the corpus to fine-tune transformer sentiment models. It is released openly under CC BY 4.0 and gives dialectal Arabic work a realistic benchmark.

Key points

• 45,000 manually labeled Algerian-dialect YouTube comments from 30+ Algerian news and media channels, collected via the YouTube Data API; spam and duplicates were removed and no personal identifiers were kept.
• Labels 0–4 (very negative → very positive) are distributed as follows: positive 16,747 (37.21%), negative 14,414 (32.03%), neutral 10,093 (22.43%), very positive 2,203 (4.90%), very negative 1,543 (3.43%).
• Record fields are text, label, last_updated, like_count, video_url and annotation_datetime; the data is tabular and ready for ML pipelines.
• About 15% of comments contain emojis. Mixing of Arabic and Latin script and French loanwords are common.
• The dataset is on Mendeley Data (DOI 10.17632/zzwg3nnhsz.2, CC BY 4.0). It was first used in Benmounah et al., Applied Sciences 13(20) 11,157 (2023), which tested Arabic normalization, emoji tokenization and URL placeholders with transformer models.

Interesting ideas

• I1. Using an ordered five-step polarity scale instead of plain positive/negative lets models learn intensity and treat errors between neighboring grades as less serious than opposite-pole errors.
• I2. The corpus keeps the raw comment text with spelling variants and mixed scripts, leaving normalization to downstream users, so the original phenomena stay available for study.
• I3. Each label carries its own annotation timestamp and source URL, which makes provenance, time-based analysis and auditing of labeling batches possible.
• I4. Disputed or context-dependent items are sent to a joint review rather than settled by one annotator. Uncertain cases get extra attention while easy ones stay cheap.
• I5. The authors treat the class skew as realistic and recommend imbalance-aware metrics or class weighting rather than rebalancing the data.

For Mova

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

• I1 (frames): prag could give evaluative speech acts a graded stance strength, and judge irony as the gap between the surface intensity and the polarity implied by context.
• I2 (tokens): crate en could store the untouched surface form next to the normalized lemma, so that informal spellings and code-switched tokens can be traced and explained rather than silently rewritten.
• I3 (learn): silver trees could record when they were agreed and from which source text, so drift or a bad teacher tier can be found and removed by date or origin.
• I4 (learn): sentences where the witness parsers disagree could be routed to a separate adjudication tier, such as a stronger LLM teacher, instead of being dropped or decided by majority vote.
• I5 (answer): test reports for the readers could give per-category scores, so that rare question types are not hidden behind a high overall accuracy.