Elamur Research⌕
☰
home / papers / 2019 / Detecting Irony in Shakespeare's Sonnets with SPARSAR

Detecting Irony in Shakespeare's Sonnets with SPARSAR

Paper · 2019 · Rodolfo Delmonte and Nicolò Busetto

Summary

The paper tackles automatic irony detection in literary text, using the 154 Shakespeare Sonnets rather than tweets. The authors first built a gold standard from literary criticism: a sonnet counts as ironic or sarcastic if at least one of the consulted critics described it that way. They then ran SPARSAR, a rule-based system for analysing and reciting English poetry. It produces phonetic, poetic-metrical and semantic representations, the semantic layer built on a modified VENSES pipeline.

The first detector flags a sonnet as ironic when two signals disagree: a sound-based mood score and lexicon-based sentiment polarity. Matched against the critics, this reached only 50%. The authors then examined the misses and used the sonnet's argumentative structure, in which the final couplet often overturns the preceding quatrains. A rule that looks at the couplet's discourse markers and negation raised accuracy to 66.88%. The work matters because it frames irony as an explainable structural reversal of mood, located in specific lines, rather than as an opaque binary label.

Key points

  • Gold standard. It is built from critics' comments (e.g., Frye, Weiser, Melchiori, Serpieri, Schoenfeldt), with a lenient rule: one critic calling a sonnet ironic is enough.
  • Sound-based mood. It averages three ratios: Contrasting Vowels, Contrasting Consonants and Contrasting Voicing. A mean above 1 counts as positive, below 1 as negative.
  • Contrast flag. A clash between sound mood and sentiment polarity (sentiment computed with propositional negation, following Taboada et al.) marks a sonnet as containing contrast.
  • Baseline result. The clash flag matched the critics on 77 sonnets, i.e. 50%.
  • Couplet rule. Of 45 missed sonnets, 26 were recovered:
    • 13 by a couplet-initial adversative, concessive or resultative marker ("but", "yet", "so", "then");
    • 13 by negation or negative items in the couplet against a mostly positive body.
  • Final accuracy and caveats. Accuracy rose to 66.88%; the conclusion states 68.88%, an internal inconsistency. Extending the rule to the last quatrain, which would cover the 19 remaining misses, was tried and abandoned.

Interesting ideas

  • I1. Irony as a clash between independent signals. The system detects irony when two separately computed channels (sound-symbolic mood vs. lexical polarity) disagree, rather than training a single classifier; this makes each decision traceable to the conflicting evidence.
  • I2. Position-sensitive reversal. Mood is evaluated separately for the body of the text and its closing unit, and a reversal triggered by an initial "but/yet/so/then" or by fresh negation is treated as the ironic cue. This exploits the fact that genres put their turn in predictable places.
  • I3. Feature vector on each main verb. Each clause's main verb gets a vector of tense, aspect, negation, modality and adverbials, which rules then use to label the proposition objective/subjective, factual/nonfactual and foreground/background. Factuality separates opinion from fact, and subjectivity counts as evidence toward irony.
  • I4. Error-driven rule induction from misses. The authors listed the system's failures against the gold standard, inspected them, and derived a new rule that fixed a clean subset. They dropped a further extension when it did not generalise.
  • I5. Three-slot topic stack for pronoun resolution. Anaphora uses Main, Secondary and Potential Topic slots, and a Potential Topic is promoted only after it is mentioned again. This gives a cheap and inspectable salience model.

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 flag incongruity when the polarity of the literal frame clashes with the polarity implied by context or by the world state, and report both sides as the explanation.
  • I2 (frames): Fables and tales also end with a turn or moral, so prag could compare the sentiment and factuality of the final sentences against the earlier world state to detect ironic or twist endings.
  • I3 (world): Storing these labels on each event would keep characters' opinions and non-factual claims out of the asserted world state, while still letting them feed irony and answer reasoning.
  • I4 (learn): Self-induced repair rules could follow the same loop, clustering failures, proposing one rule, keeping it only if it fixes a coherent group, and discarding it otherwise.
  • I5 (frames): Coreference could use this promotion-by-reiteration rule as a transparent baseline and fallback for resolving free pronouns in short narratives.

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 5 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works Literature and Literary Theory · 340 works

🧭 Topics 2 English · 1,373 works Sentiment & emotion · 608 works

Cite

Cite as Rodolfo Delmonte and Nicolò Busetto. 2019. Detecting Irony in Shakespeare's Sonnets with SPARSAR. Proceedings of the Sixth Italian Conference on Computational Linguistics (CLiC-it 2019).

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

@inproceedings{delmonte2019detecting,
  title = {{Detecting Irony in Shakespeare's Sonnets with SPARSAR}},
  author = {Rodolfo Delmonte and Nicolò Busetto},
  booktitle = {Proceedings of the Sixth Italian Conference on Computational Linguistics (CLiC-it 2019)},
  year = {2019},
  url = {https://aclanthology.org/2019.clicit-1.23/}
}

References

24 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Salvatore Attardo. 1994. Linguistic Theories of Humor. Mouton de Gruyter, Berlin New York. find
  2. Salvatore Attardo. 2000. Irony as relevant inappropriateness. Journal of Pragmatics, 84(32). ↗
  3. Ciprian Bacalu and Rodolfo Delmonte. 1999. Prosodic modeling for syllable structures from the vesd - venice english syllable database. In Aspetti Computazionale in Fonetica, Linguistica e Didattica delle Lingue: Modelli e Algoritmi - Atti 9 Convegno GFS-AIA, pages 147–160, Venezia. GFS-AIA. find
  4. Joan Bresnan, editor. 1982. The Mental Representation of Grammatical Relations. The MIT Press, Cambridge MA. find
  5. Joan Bresnan. 2001. Lexical-Functional Syntax. Blackwell Publishing, Oxford. find
  6. Dario Calimani. 2009. William Shakespeare, I sonetti della menzogna. Carrocci, Roma. find
  7. P. Carvalho, L. Sarmento, M. Silva, and E. de Oliveira. 2009. Clues for detecting irony in user-generated contents: oh...!! it’s so easy;-). In Proceeding of the 1st international CIKM workshop on Topic-sentiment analysis for mass opinion, pages 53–56. ACM. find
  8. Rodolfo Delmonte, Sara Tonelli, Marco Aldo Piccol-ino Boniforti, Antonella Bristot, and Emanuele Pianta. 2005. Venses – a linguistically-based system for semantic evaluation. In Machine Learning Challenges, pages 344–371, Berlin. Springer. find
  9. Rodolfo Delmonte. 2016. Exploring shakespeare’s sonnets with sparsar. volume 4, pages 61–95. find
  10. R.L. Eagle. 1916. New light on the enigmas of Shakespeare’s Sonnets. John Long Limited, London. find
  11. Ivan Fonagy. 1971. The functions of vocal style. In Seymour Chatman, editor, Literary Style: A Symposium, pages 159–174. Oxford UP, London. find
  12. Northrop Frye. 1957. Anatomy of Criticism: Four Essays. Princeton University Press. find
  13. Baayen R. H., R. Piepenbrock, and L. Gulikers. 1995. The CELEX Lexical Database (CD-ROM). Linguistic Data Consortium. find
  14. Justine Kao and Dan Jurafsky. 2012. A computational analysis of style, affect, and imagery in contemporary poetry. In Proceedings of NAACL Workshop on Computational Linguistics for Literature, pages 8– 17, Stroudsburg, PA, USA. ACL. find
  15. S.-M. Kim and E. Hovy. 2004. Determining the sentiment of opinions. In Proceedings of the 20th international conference on computational linguistics -COLING, pages 1367–1373, Stroudsburg, PA, USA. ACL. our card
  16. Maria Antonietta Marelli. 2015. William Shakespeare, I Sonetti – con testo a fronte. Garzanti, Milano. find
  17. Giorgio Melchiori. 1971. Adriatica Editrice, Bari. Shakespeare’s Sonnets. find
  18. Antonio Reyes and Paolo Rosso. 2013. On the difficulty of automatically detecting irony: beyond a simple case of negation. Knowledge and Information Systems, 40:595–614. ↗
  19. Antonio Reyes, Paolo Rosso, and Davide Buscaldi. 2012. From humor recognition to irony detection: The figurative language of social media. Data & Knowledge Engineering. ↗
  20. Michael Schoenfeldt. 2010. Cambridge introduction to Shakespeare’s poetry. Cambridge University Press, Cambridge. find
  21. Alessandro Serpieri. 2002. Polifonia Shakespeariana. Bulzoni, Roma. find
  22. M. Taboada, J. Brooke, M. Tofiloski, K. Voll, and M. Stede. 2011. Lexicon-based methods for sentiment analysis. Computational Linguistics, 37(2):267–307. ↗ our card
  23. Reuven Tsur. 1992. What Makes Sound Patterns Expressive: The Poetic Mode of Speech-Perception. Duke UP, Durham N.C. find
  24. Cynthia Van Hee, Els Lefever, and Véronique Hoste. 2018. Semeval-2018 task 3: Irony detection in english tweets.). In Proceedings of The 12th International Workshop on Semantic Evaluation, pages 39– 50. ACL. David K. Weiser. The ’sonnets’. 84(4):456–469. 1983. Shakespearean irony: Neuphilologische Mitteilungen, David K. Weiser. 1987. Mind in Character – Shakespeare’s Speaker in the Sonnets. The University of Missouri Press. APPENDIX List of couplets and quatrains from sonnets which contain a discourse marker for reverted logical structure our card

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/delmonte-2019-detecting-irony-shakespeare-s-sonnets/