📄 TUNIZI: a Tunisian Arabizi sentiment analysis Dataset

Summary

Tunisians writing online often use Arabizi, a romanized form of their dialect in which digits stand for Arabic sounds that the Latin alphabet lacks. Before this work, no annotated Arabizi resource existed for this dialect. Earlier surveys had found only Arabic-script Tunisian corpora, and a parallel-corpus study had shown that Tunisian shares the least vocabulary with other Arabic dialects. The authors built TUNIZI, a sentiment corpus of YouTube comments taken from videos across many genres. They used scripted scraping, automatic removal of French and English comments, and a manual pass to discard any remaining non-Arabizi text. Links, emoji and punctuation were stripped. Five native speakers labelled each comment as positive or negative, and two of them then reviewed the labels. The result is a corpus of about 9.2k comments with a near-even polarity split, released publicly on GitHub. The paper reports no classification baselines. Its value is as a first labelled resource for a low-resource, non-standard North African variety.

Key points

• TUNIZI contains 9,210 comments: 4,838 negative and 4,372 positive (roughly 53% / 47%), with 79,862 tokens and 22,850 distinct words.
• The comments come from YouTube videos on sports, politics, comedy, TV shows and series, arts and Tunisian music, chosen to cover varied writers and styles.
• Filtering had two steps: a script removed French and English comments (the most common languages among Tunisians), and a manual check removed the remaining non-Arabizi items.
• Five annotators (three men, two women, all at Master/PhD level) split the labelling equally, and the two women then verified the annotation.
• The authors cite Bouamor et al. (2014), whose Overlap Coefficient showed Tunisian is the least lexically similar of five dialects. They propose similar datasets for Algerian and Moroccan Arabizi as future work.

Interesting ideas

• I1. Collecting from deliberately varied content genres (sports, politics, music, comedy) spreads the corpus over many registers and writer profiles, so it is not skewed toward one domain. — For Mova (learn): When gathering raw English text for silver data agreed by witness parsers, stratify sources by genre so the training signal does not overfit one register.
• I2. Cheap automatic filtering removes the most likely contaminating languages first, and a human pass catches what slips through. This keeps manual effort small while still producing a clean corpus. — For Mova (learn): Run a fast automatic filter for non-English or malformed input before silver-data agreement, then send only the residue to an LLM teacher tier for a final keep/drop decision.
• I3. Labelling is split across several annotators, and a smaller subset then reviews the whole set, giving a simple two-tier quality check. — For Mova (learn): Mirror this with a first-pass teacher producing frame or role labels and a stronger reviewer tier auditing them, logging disagreements as candidates for self-induced repair rules.
• I4. A lexical overlap coefficient quantifies how distant one language variety is from its relatives, which shows how much transfer to expect before investing in new data. — For Mova (tree): Measure vocabulary overlap between a target text domain and each available treebank to decide which treebank or UD dialect conversion to rely on for gparse fine-tuning.
• I5. In Arabizi, digits embedded in words encode sounds, so a mixed letter-digit token is a word, not a number. — For Mova (tokens): Have the en tokenizer and tagger treat alphanumeric blends (usernames, leetspeak, codes) as single word tokens rather than splitting off numerals, so the tree and the world's measures are not polluted by spurious numbers.