Elamur Research⌕
☰
home / papers / 2016 / Sarcasm Detection : Building a Contextual Hierarchy

Sarcasm Detection : Building a Contextual Hierarchy

Paper · 2016 · Taradheesh Bali and Navjyoti Singh

Summary

The paper argues that sarcasm cannot be recognised reliably from the words of a single utterance. It depends on what the speaker and the listener already know and believe about the topic. After reviewing Gricean, echoic and pretense theories, the authors model sarcasm as an exchange between two agents, using a hierarchy of shared knowledge between them. A tweet counts as contradicting a belief when its sentiment toward a topic clashes with the belief attributed to the audience. The source of that belief depends on the audience type:

  • Earlier exchanges: if the speaker and listener have discussed the topic before, the belief comes from those exchanges.
  • No shared history: otherwise, it comes from the listener's own posts and from recent tweets by everyone else, which stand in for social norms.

They built a balanced Twitter corpus of reply conversations using #sarcasm/#sarcastic markers and trained an l2-regularised logistic regression on lexical, pragmatic and contextual features. Adding the binary belief-contradiction feature raised accuracy from 75.8% to 78.7%. The work matters because it makes the speaker–listener relation an explicit, computable part of context, even when the two have never interacted.

Key points

  • Data: 2000 conversational instances. The 1000 sarcastic ones are split evenly between the two audience cases (500 each), and there are 1000 non-sarcastic ones. Negatives were checked by hand to remove implicit sarcasm, and a Reference Table keeps users, tweets and retweets from overlapping with the positive set.
  • Belief sets: each belief set holds at most 10 tweets about elements of the hierarchy H. For an audience that has interacted before (L1), these are tweets the speaker sent to the listener. Otherwise (L2), they are the listener's own tweets plus recent tweets from the wider world.
  • Feature groups:
    • Lexical: word unigrams/bigrams, Brown clusters, CMU TweetNLP POS tags.
    • Pragmatic: capital-letter count, emoticons and expressions, top-100 tf-idf sarcastic words.
    • Contextual: belief contradiction.
  • Pipeline:
    • Tokens are cleaned with the English Normalizer of Sharma et al. (2016).
    • Stanford's Sentiment Analyzer scores beliefs on a scale from -1 to 1.
    • Each belief set's score is the average over its tweets.
  • Results: a majority-class baseline gives 48%. Under 10-fold cross-validation (8 folds train / 1 tune / 1 test), accuracy is 71.2% for lexical features, 75.8% for lexical+pragmatic, and 78.7% with the contextual feature added.

Interesting ideas

  • I1. The model asks whether there is any shared history with this particular listener. If not, it falls back to public opinion as a stand-in norm, so context degrades gracefully instead of disappearing.
  • I2. Sarcasm is treated as a clash between the stance in the current utterance and the stance the audience is assumed to hold, rather than as a property of the words themselves. That makes each decision traceable to the specific belief it contradicts.
  • I3. The audience's view is split into two sources: what this person personally expressed (their bias) and what the crowd says (the norm).
  • I4. Negatives are cross-checked against a reference table so that no user, tweet or retweet leaks between classes. This is a simple guard against inflated scores.
  • I5. The authors suggest scoring sentiment per topic inside an utterance instead of over the whole utterance, so the contrast is found at a sub-sentence level.

For Mova

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

  • I1 (frames): the irony detector in prag could first look for prior statements by the same characters in the text, and only then fall back to default expectations from compiled knowledge.
  • I2 (frames): an irony frame could record a pointer to the contradicted belief, so the explanation shows which earlier state or norm was inverted.
  • I3 (world): the world model could keep a character's stated attitudes apart from generic norms, letting readers reason about whose expectation an utterance violates.
  • I4 (learn): silver-data splits and negative controls could enforce the same no-shared-source rule across train and test documents.
  • I5 (frames): stance could be attached to individual frame arguments in the UD-based frames, so a contradiction is localised to one target entity.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

📚 Shelves 4 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Psychology · 798 works Social Psychology · 129 works

🧭 Topics 1 Sentiment & emotion · 608 works

Cite

Cite as Taradheesh Bali and Navjyoti Singh. 2016. Sarcasm Detection : Building a Contextual Hierarchy. Proceedings of the Workshop on Computational Modeling of People's Opinions, Personality, and Emotions in Social Media (PEOPLES).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{bali2016sarcasm,
  title = {{Sarcasm Detection : Building a Contextual Hierarchy}},
  author = {Taradheesh Bali and Navjyoti Singh},
  booktitle = {Proceedings of the Workshop on Computational Modeling of People's Opinions, Personality, and Emotions in Social Media (PEOPLES)},
  year = {2016},
  url = {https://aclanthology.org/W16-4313/}
}

References

25 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. David Bamman and Noah A Smith. 2015. Contextualized sarcasm detection on twitter. In Ninth International AAAI Conference on Web and Social Media. find
  2. Francesco Barbieri, Horacio Saggion, and Francesco Ronzano. 2014. Modelling sarcasm in twitter, a novel approach. ACL 2014, page 50. our card
  3. Elisabeth Camp. 2012. Sarcasm, pretense, and the semantics/pragmatics distinction. Noûs, 46(4):587–634. ↗
  4. Herbert H Clark and Richard J Gerrig. 1984. On the pretense theory of irony. Journal of Experimental Psychology: General, 113(1):121–126. ↗
  5. HH Clark. 1996. Using language cambridge university press cambridge. find
  6. Gregory Currie. 2006. Why irony is pretence. The architecture of the imagination, pages 111–33. ↗
  7. Dmitry Davidov, Oren Tsur, and Ari Rappoport. 2010. Semi-supervised recognition of sarcastic sentences in twitter and amazon. In Proceedings of the fourteenth conference on computational natural language learning, pages 107–116. Association for Computational Linguistics. our card
  8. Henry Watson Fowler. 1965. Modern english usage (revised by sir ernest gowers). find
  9. Aniruddha Ghosh, Guofu Li, and Tony Veale. 2015. Semeval-2015 task 11: Sentiment analysis of figurative language in twitter. our card
  10. Kevin Gimpel, Nathan Schneider, Brendan O’Connor, Dipanjan Das, Daniel Mills, Jacob Eisenstein, Michael Heilman, Dani Yogatama, Jeffrey Flanigan, and Noah A Smith. 2011. Part-of-speech tagging for twitter: Annotation, features, and experiments. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: short papers-Volume 2, pages 42–47. Association for Computational Linguistics. ↗ our card
  11. Roberto González-Ibánez, Smaranda Muresan, and Nina Wacholder. 2011. Identifying sarcasm in twitter: a closer look. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: short papers-Volume 2, pages 581–586. Association for Computational Linguistics. our card
  12. Herbert P Grice. 1970. Logic and conversation. na. find
  13. H Paul Grice. 1978. Further notes on logic and conversation. 1978, 1:13–128. ↗
  14. Aditya Joshi, Vinita Sharma, and Pushpak Bhattacharyya. 2015. Harnessing context incongruity for sarcasm detection. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, volume 2, pages 757–762. our card
  15. Sachi Kumon-Nakamura, Sam Glucksberg, and Mary Brown. 1995. How about another piece of pie: The allu-sional pretense theory of discourse irony. Journal of Experimental Psychology: General, 124(1):3. find
  16. Benny PH Lee. 2001. Mutual knowledge, background knowledge and shared beliefs: Their roles in establishing common ground. Journal of pragmatics, 33(1):21–44. ↗
  17. Mihaela Popa-Wyatt. 2014. Pretence and echo: towards an integrated account of verbal irony. International Review of Pragmatics, 6(1):127–168. ↗
  18. Ellen Riloff, Ashequl Qadir, Prafulla Surve, Lalindra De Silva, Nathan Gilbert, and Ruihong Huang. 2013. Sarcasm as contrast between a positive sentiment and negative situation. In EMNLP, volume 13, pages 704–714. our card
  19. Arnav Sharma, Sakshi Gupta, Raveesh Motlani, Piyush Bansal, Manish Srivastava, Radhika Mamidi, and Dipti M Sharma. 2016. Shallow parsing pipeline for hindi-english code-mixed social media text. arXiv preprint arXiv:1604.03136. ↗
  20. Richard Socher, Alex Perelygin, Jean Y Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the conference on empirical methods in natural language processing (EMNLP), volume 1631, page 1642. Citeseer. find
  21. Dan Sperber. 1984. Verbal irony: Pretense or echoic mention? ↗
  22. Emilio Sulis, Delia Irazú Hernández Farı́as, Paolo Rosso, Viviana Patti, and Giancarlo Ruffo. 2016. Figurative messages and affect in twitter: differences between# irony,# sarcasm and# not. Knowledge-Based Systems. ↗
  23. Oren Tsur, Dmitry Davidov, and Ari Rappoport. 2010. Icwsm-a great catchy name: Semi-supervised recognition of sarcastic sentences in online product reviews. In ICWSM. find
  24. Kendall L Walton. 1990. Mimesis as make-believe: On the foundations of the representational arts. Harvard University Press. ↗
  25. Deirdre Wilson. 2006. The pragmatics of verbal irony: Echo or pretence? Lingua, 116(10):1722–1743. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bali-2016-sarcasm-detection-building-contextual-hierarchy/