📄 ArSarcasMoji Dataset: The Emoji Sentiment Roles in Arabic Ironic Contexts

Summary

Arabic irony and sarcasm research has usually removed emoji during preprocessing, on the assumption that they complicate an already morphologically rich language. This paper argues the opposite and releases ArSarcasMoji, a corpus of 24,630 short Arabic social-media texts that contain emoji. Of these, 4,320 are labelled ironic (sarcastic or humorous) and 20,310 are not.

The authors pooled emoji-bearing tweets from 15 existing Arabic datasets and made a twin copy of each tweet with the emoji stripped out. They kept only the items on which five sentiment classifiers agreed. They then used an earlier emoji-sentiment-role model to work out what each emoji does to the polarity of its text, for example strengthening it, softening it or flipping it. A tweet is marked ironic when it contains an ironic emoji pattern grown from 23 seed emoji and that pattern plays a particular sentiment role.

The analysis reports how sarcasm and humour are spread across these roles. One salient finding is that emoji which soften negative text are frequently sarcastic. The resource gives Arabic irony work a way to treat emoji as signal rather than noise.

Key points

• Corpus size and sources: 15 Arabic datasets (AraSenCorpus, ArSarcasm, ASAD, ArSAS, SemEval tasks and others) supplied 367,120 emoji-containing tweets. Filtering reduced these to the final 24,630 texts, of which about 17.5% are ironic.
• Silver sentiment labels: both the Emoji-Text and Plain-Text versions were labelled by Mazajak, CAMeL-Tools (AraBERT and mBERT), ASAD and a lexicon-based method. A tweet was kept only if all of them agreed.
• Ironic emoji patterns: the 23 seed emoji were combined with emoji from the 1,272 seen in the raw data, leaving out flags, some nature emoji and most hearts. This produced 15,101 ironic emoji patterns: 8,990 positive, 543 neutral and 5,568 negative.
• Sentiment-role inventory: Emphasizer, Indicator, Mitigator, Reverser and Trigger, each for positive or negative sentiment, plus a Neutral Effect.
• Distributions:
  • Texts split into 54.64% positive, 44.63% negative and 0.73% neutral.
  • Positivity Emphasizer accounts for 53.26% of emoji in positive texts, with 5.39% of those humorous.
  • Negativity Emphasizer accounts for 36.23%, with 24.57% sarcastic.
  • Negativity Mitigator accounts for 7.73%, and 68.99% of those are sarcastic.
  • Trigger and Reverser roles are very rare, which the authors list as a limitation.

Interesting ideas

• I1. Paired texts with and without the cue: the role of a non-verbal cue is defined by comparing sentiment on the text with the cue against the same text without it, which keeps the cue's contribution measurable and transparent. — For Mova (frames): prag could compare the literal clause polarity with and without markers such as emoji, "yeah, right" or scare quotes, and record the shift as an explicit irony feature.
• I2. Small role taxonomy for modifiers: a cue that emphasizes, mitigates, reverses, triggers or indicates sentiment gives a compositional, explainable label instead of a single opaque irony score. — For Mova (frames): the irony and implicature module could tag each evaluative marker with one of these roles, so an answer can state why a sentence was read as sarcastic.
• I3. Unanimity across heterogeneous labellers: only items on which five different classifiers, including neural and lexicon-based ones, agreed were kept, trading coverage for label reliability. — For Mova (learn): silver sentiment or speech-act labels could require full agreement among diverse witnesses, mirroring the multi-parser silver trees.
• I4. Growing a marker lexicon from seeds: a handful of irony emoji drawn from prior studies was expanded into a large pattern inventory using correlations within the corpus. — For Mova (world): compiled knowledge could grow an irony and hedge marker list from a few hand-picked seeds by co-occurrence, while keeping the seeds as provenance for each entry.
• I5. Reporting rare categories openly: the authors state that some roles have almost no examples rather than averaging them away, which marks where conclusions are weak. — For Mova (learn): evaluation reports could break results down per pragmatic role and flag classes that are too sparse to trust.