📄 "With 1 follower I must be AWESOME :P". Exploring the role of irony markers in irony recognition

Summary

The paper asks which surface cues help a classifier separate ironic from non-ironic social media posts. It does not aim for a state-of-the-art detector. It builds on the theory of irony markers from Attardo and Burgers. These markers signal ironic intent, but deleting them does not destroy the irony itself. The authors group markers into three families: tropes (metaphor, hyperbole, rhetorical questions), morpho-syntactic cues (exclamations, tag questions, interjections) and typographic cues (capitals, quotes, punctuation, hashtags, emoticons, emoji). They encode these as features and train a linear SVM only on them. Data are balanced, self-labelled sets: tweets with irony/sarcasm hashtags and Reddit posts marked "/s". Ablation, SVM weights and frequency statistics all point the same way. On Twitter, emoji and emoticons are the most telling cues. On Reddit, morpho-syntactic cues such as interjections and tag questions matter most, and typography is nearly absent. Marker use also varies by topic, rising in political and religious subreddits. The work shows that useful irony cues depend on the platform and genre.

Key points

• Data: 350K English tweets labelled via #irony/#sarcasm/#sarcastic, with retweets, duplicates and non-final hashtags removed. There are 50K Reddit replies of at least two sentences from the Khodak, Saunshi and Vodrahalli (2018) corpus. Both sets are balanced and split 80/10/10.
• Resources used:
  • adjective/noun metaphor lists: 884 pairs from Tsvetkov et al. 2014 and 8,600 from Gutiérrez et al. 2016;
  • the Rei et al. 2017 metaphor detector, keeping candidates with precision ≥ 0.75;
  • MPQA "strong subjective" words as hyperboles;
  • about 250 interjections, over 100 emoticons and a 1,400-emoji library.
• Twitter, all features: irony F1 is 71.75 and non-irony F1 is 66.97. Without tropes (mainly hyperbole), irony F1 falls to 56.18. Removing typography hurts irony precision most.
• Reddit, all features: irony F1 is 58.35 and non-irony F1 is 70.35. Without morpho-syntactic markers these fall to 53.49 and 60.16. Removing typography changes almost nothing (58.39 / 70.42).
• Frequency analysis:
  • emoji and emoticons are common on Twitter but virtually unused on Reddit;
  • exclamations and rhetorical questions are more frequent on Reddit;
  • differences between marker types are significant (t-test, p ≤ 0.005);
  • politics and religion subreddits use more hyperbole, rhetorical questions, exclamations and punctuation than technology and sports.

Interesting ideas

• I1. The theory separates removable markers, which only signal irony, from essential factors, which make up the irony. A detector built on markers alone therefore measures signalling, not the irony itself.
• I2. A question that sits mid-comment, followed by more text, is treated as probably rhetorical rather than a real request for information.
• I3. The polarity of a tag question carries signal: negative tags ("isn't it?") lean ironic, while positive tags lean literal on Reddit.
• I4. Feature families are ablated as whole groups and the learned weights are inspected, so the result says which kind of cue matters, not just a single score.
• I5. Compound hashtags are split into words before sentiment lookup, so packed tokens like #funnight still yield usable lexical evidence.

For Mova

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

• I1 (frames): The irony module in prag could store markers as separate, citable evidence next to the core cue, such as a clash between literal sentiment and the described situation, so the explanation shows which kind of evidence fired.
• I2 (frames): The speech-act tagger could use sentence position within a turn as a cheap explainable rule for downgrading questions from "request for information" to "rhetorical assertion".
• I3 (frames): Since UD trees expose the tag clause and its negation, prag could read tag polarity directly from the tree as an irony or implicature feature.
• I4 (learn): Mova's evaluation could ablate whole rule or evidence families per genre, giving interpretable reports and acting as a negative control on which components actually contribute.
• I5 (tokens): The en tokenizer could split hashtags and handles into words and lemmatize the parts, letting downstream lexicons and frames see their content.