📄 Exploring the Impact of Pragmatic Phenomena on Irony Detection in Tweets: A Multilingual Corpus Study

Summary

Irony detectors for Twitter mostly rely on cues inside the message, but much irony depends on knowledge the reader brings from outside. This paper asks four questions. Do the irony categories that linguists described for literary texts also appear in tweets? Do they carry surface markers? How do categories and markers co-occur? Do French, English and Italian favour different categories? The authors propose a three-layer annotation scheme. The first layer records how irony is triggered: explicitly, through a contradiction between two lexicalized propositions, or implicitly, through a clash with an unstated pragmatic context. The second layer records the irony category, and the third records the linguistic markers present. They apply the scheme to new French and English tweet collections on nine media topics and to existing Italian data, then analyse the results qualitatively and quantitatively. The main finding is that implicitly triggered irony, which needs external knowledge, is the hardest case for future systems. The released corpus is a fine-grained multilingual resource and has already supported irony detection shared tasks.

Key points

• Corpora:
  • French (F): 2,073 ironic vs. 16,179 non-ironic tweets.
  • English (E): 5,173 ironic vs. 6,116 non-ironic tweets.
  • Italian (I): 806 ironic tweets from Sentipolc plus 2,273 from TW-SPINO, with 5,642 non-ironic Sentipolc tweets.
  • The corpus is freely available at github.com/IronyAndTweets.
• Data collection: French and English tweets were gathered through the Twitter API using keywords for nine topics, such as politics, sport and the Arab Spring. Ironic examples were identified by hashtags like #irony and #sarcasm. Italian needed existing annotated data because Italian users rarely use stable irony hashtags.
• Scheme layers:
  • Irony activation (EXPLICIT vs. IMPLICIT).
  • Irony categories adapted from the literature, such as analogy, hyperbole/exaggeration, euphemism, rhetorical question, oxymoron, false assertion and context shift.
  • Markers, such as negation, punctuation, intensifiers, URLs and named entities.
• Pilot study: The literary categories were first checked on a pilot sample of 150 ironic tweets.
• Shared tasks: The French data was used for DEFT@TALN2017 and the Italian data for SENTIPOLC@Evalita (2014 and 2016). In the 2016 edition, 13 teams submitted 57 runs. Irony detection proved difficult, partly because the data is unbalanced: only about 1 in 7 training tweets is ironic.