📄 Exploratory Arabic Offensive Language Dataset Analysis

Summary

The authors ask what is actually inside the openly released Arabic offensive-language datasets, where those datasets fall short, and what new resources should be built. They gathered nine freely available corpora covering hate speech, obscene and abusive text in several dialects. Each corpus was reduced to a uniform layout of text plus label, with one raw text file per class. They then ran an exploratory data analysis with three parts. The statistical part covered word and stop-word frequencies, distinctive words per class found with Voyant, tokens per post and characters per token. The textual part used word clouds. The contextual part looked at sentiment predicted by Mazajak, emoji usage and punctuation.

Most datasets have short posts, averaging 15 tokens or fewer. Most are also imbalanced, with only a small share of offensive examples. The top offensive vocabulary barely overlaps between corpora, which the authors attribute to dialect and sub-culture differences. They therefore argue that no single dataset is enough for training a detector. Instead, knowledge should be pooled across all of them.

Key points

• The nine datasets are:
  • Aljazeera.net Deleted Comments and Egyptian Tweets (Mubarak et al., 2017)
  • YouTube Comments (Alakrot et al., 2018)
  • Religious Hate Speech (Albadi et al., 2018)
  • L-HSAB (Levantine)
  • T-HSAB (Tunisian)
  • MPOLD (Chowdhury et al., 2020)
  • OSACT4/OffensEval (Mubarak et al., 2020)
  • Multi-Platform Hate Speech (Omar et al., 2020)
• The Aljazeera dataset has 31,692 comments: 5,653 clean, 533 obscene and 25,506 offensive, with only 8 duplicates. Its offensive comments are the longest of the three classes.
• Stop words were removed using the NLTK Arabic list combined with the list from Albadi et al. The definite-article prefix was stripped except where it belongs to the word itself.
• POS tagging was skipped because available tools cannot handle unstructured dialectal user text.
• Cross-dataset findings:
  • 😂 is the most frequent emoji in both offensive and non-offensive samples.
  • Punctuation shows no consistent pattern.
  • The vocative "يا / you" frequently ranks among the top stop words in offensive classes.
  • Names of famous people, such as Kadhim and Ahlam in the YouTube corpus, appear among the frequent tokens.

Interesting ideas

• I1. Count stop words separately for each class, and treat a "function word" that clusters in one class as a real feature instead of discarding it. Fixed stoplists can silently delete signal. — For Mova (tokens): audit any stoplist or closed-class filtering in crate en by comparing frequencies across text genres or labels before hard-coding what counts as noise.
• I2. The vocative particle "you" being tied to offensive posts suggests that direct address of a hearer is a useful cue for hostile speech acts. — For Mova (frames): in crate prag, use vocative/addressee dependencies from the UD tree as explicit evidence when classifying a speech act as an insult or attack.
• I3. Compare the top distinctive tokens across corpora built for the same task. Low overlap exposes dialect or domain shift that a single-corpus score would hide. — For Mova (learn): profile silver data from different sources by lexical overlap per label, and use low-overlap sources as a held-out negative control for generalization.
• I4. Analyze raw, uncleaned text in a minimal uniform format (text, label, one file per class) before any preprocessing, so the analysis does not erase phenomena like emojis or punctuation. — For Mova (learn): keep an untouched raw copy of every ingested corpus next to the normalized one, so repair rules can be checked against what the text originally contained.
• I5. Very short posts (around 15 tokens or fewer) make offense hard to separate from sarcasm, because the needed context is missing. — For Mova (frames): when irony or implicature detection has too little context, return an explicit "insufficient context" verdict instead of a forced label.