---
title: "A Neural Approach to Irony Generation — Zhu 2019"
description: "The paper presents irony generation as an unsupervised style-transfer task. 26 references, cited by 4 papers here."
authors: ["Mengdi Zhu", "Zhiwei Yu", "Xiaojun Wan"]
year: 2019
venue: "arXiv:1909.06200"
source: "https://arxiv.org/abs/1909.06200"
doi: "10.48550/arxiv.1909.06200"
arxiv: "1909.06200"
licence: "arXiv non-exclusive distribution licence (read only, no redistribution)"
licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/zhu-2019-neural-approach-irony-generation/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/zhu-2019-neural-approach-irony-generation/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/) / A Neural Approach to Irony Generation

# A Neural Approach to Irony Generation

**Authors:** Mengdi Zhu, Zhiwei Yu and [Xiaojun Wan](https://md.research.elamur.ai/free/en/authors/xiaojun-wan/) · **Year:** 2019 · **Venue:** arXiv:1909.06200 · **Licence:** [arXiv non-exclusive distribution licence (read only, no redistribution)](http://arxiv.org/licenses/nonexclusive-distrib/1.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/1909.06200) · [PDF](https://arxiv.org/pdf/1909.06200) · [DOI 10.48550/arxiv.1909.06200](https://doi.org/10.48550/arxiv.1909.06200) · [arXiv 1909.06200](https://arxiv.org/abs/1909.06200) · [OpenAlex](https://openalex.org/W2972339230)

**Citations:** 5 citations (OpenAlex, 2026-10-14) · [4](https://md.research.elamur.ai/free/en/papers/zhu-2019-neural-approach-irony-generation/cited-by/) in our library, of the 141,840 papers whose bibliographies we read

## Summary
The paper presents irony generation as an unsupervised style-transfer task. A non-ironic sentence is rewritten as an ironic one while keeping its sentiment polarity and its sentiment-unrelated content. An example is turning a sentence about hating ruined plans into one that claims to love them. The authors restrict the scope to polarity-contrast verbal irony and set aside the other two irony types. No parallel data existed, so they crawled over 2M English tweets and cleaned them. An LSTM classifier trained on SemEval-2018 Task 3 then split the tweets into ironic and non-ironic pools. The model uses two encoders and two decoders, one pair per style, with some shared layers. It is first pre-trained with a denoising auto-encoder and back-translation, then fine-tuned with reinforcement learning using two rewards: one for irony and one for sentiment preservation. The authors report good irony accuracy with well-kept sentiment and content compared with style-transfer baselines. They also test the reverse direction, ironic to non-ironic. The work matters because it supplies the first large irony dataset and a benchmark for a task that had been studied almost only from the detection side.

## Key points
- **Dataset construction**
  - Pipeline: tweets from 2012–2018 → removal of retweets and non-English text (langdetect) → normalization with Ekphrasis → abbreviation expansion via a webopedia dictionary → length filter of 10–40 tokens → rare-word filter. This left 662,530 sentences.
  - The LSTM classifier labeled 262,755 sentences ironic and 399,775 non-ironic.
  - Non-ironic sentences were then reduced to non-questions containing a strongly polar word (TextBlob score magnitude above 0.5). The final count is 102,330, although the introduction says 112,330.
- **Model design:** style-specific encoder/decoder pairs with shared layers to force a common latent space. Pre-training uses DAE plus back-translation, followed by RL with separate irony and sentiment rewards.
- **Baselines:** BackTrans, Unpaired, CrossAlign, CPTG, DualRL.
- **Reverse direction (ironic → non-ironic):**
  - The model has the lowest Senti Delta (0.4976).
  - Its BLEU is 62.92, with G2 at 57.33 and H2 at 56.64. DualRL is higher on these three scores (BLEU 74.31).
  - In human evaluation, the model ranks second in sentiment and content preservation. Both it and DualRL score poorly on irony, partly because the baselines' irrelevant outputs were easily judged non-ironic.
- **Error types analysed:**
  - No change to the input sentence, because content rewards dominate.
  - Word repetition, because the model chases high classifier scores.
  - Edits to sentiment-irrelevant words, such as "friday" or "fifa" instead of the polar word.

## Interesting ideas
- The task is defined as a polarity flip of the literal evaluative word while the intended sentiment stays fixed, so irony becomes a detectable clash between surface polarity and situation polarity.
- A small gold set (SemEval-2018) is used to train a deliberately simple classifier, which then labels a huge noisy pool; this avoids overfitting a complex model to scarce data.
- Non-ironic candidates are filtered to sentences with an explicit, strong-polarity word and no questions, keeping only inputs where the transformation is well-defined.
- Content preservation and style change pull against each other, so separate rewards (style, sentiment, content via back-translation) are balanced explicitly; the error analysis names the failure modes of each imbalance (copying the input vs. word repetition).
- Repetition is curbed only at decoding time by down-weighting already-emitted words; applying the same penalty during training hurt results.

## For Mova

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

- **I1** (frames): `prag` could flag irony when a strongly positive evaluative predicate is attached to a situation that world knowledge marks as negative (e.g. "love" + "plans get ruined"), then flip the speaker's intended stance in the frame.
- **I2** (learn): this is a cheap way to bootstrap silver irony or speech-act labels, which should then be checked by other witnesses and negative controls before training on them.
- **I3** (learn): when building training pairs for any transformation, first prune inputs where the target operation has no anchor, so that evaluation measures the operation rather than noise.
- **I4** (back): a graph-to-text generator can score outputs on independent axes (meaning-graph match, target register, stance preserved) and log which axis caused a rejection, keeping generation explainable.
- **I5** (back): constraints like these can be kept as an explicit, inspectable decoding rule instead of baked into training, so the realizer stays debuggable.

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

## Institutions

- [Peking University](https://md.research.elamur.ai/free/en/universities/peking-university/) — University, Beijing, China · 100 works · 35,618 citations

📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works

🧭 Topics (1): [😊 Sentiment & emotion](https://md.research.elamur.ai/free/en/topics/sentiment-emotion/) · 608 works

## Cite

Cite as: Mengdi Zhu, Zhiwei Yu and Xiaojun Wan. 2019. A Neural Approach to Irony Generation. arXiv:1909.06200.

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

```bibtex
@misc{zhu2019neural,
  title = {{A Neural Approach to Irony Generation}},
  author = {Mengdi Zhu, Zhiwei Yu and Xiaojun Wan},
  year = {2019},
  doi = {10.48550/arxiv.1909.06200},
  eprint = {1909.06200},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.1909.06200}
}
```

- [← previous What A Sunny Day ☔: Toward Emoji-Sensitive Irony Detection](https://md.research.elamur.ai/free/en/papers/hayati-2019-what-sunny-day-emoji-sensitive/)
- [next → Acceptability of at-issue co-speech gestures under contrastive focus](https://md.research.elamur.ai/free/en/papers/esipova-2019-acceptability-issue-co-speech-gestures/)
- [↑ back to papers of 2019](https://md.research.elamur.ai/free/en/papers/2019/page/6/)

## References

26 references, 16 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Aniruddha Ghosh, Guofu Li, Tony Veale, Paolo Rosso, Ekaterina Shutova, John A. Barnden, and Antonio Reyes. [Semeval-2015 task 11: Sentiment analysis of figurative language in twitter](https://md.research.elamur.ai/free/en/papers/ghosh-2015-semeval-2015-task-11-sentiment/). In SemEval@NAACL-HLT, 2015.
2. Jihen Karoui, Farah Benamara, Véronique Moriceau, Viviana Patti, Cristina Bosco, and Nathalie Aussenac-Gilles. [Exploring the impact of pragmatic phenomena on irony detection in tweets: A multilingual corpus study](https://md.research.elamur.ai/free/en/papers/karoui-2017-exploring-impact-pragmatic-phenomena-irony/). In EACL (1), 2017.
3. Sara Rosenthal, Alan Ritter, Preslav Nakov, and Veselin Stoyanov. [Semeval-2014 task 9: Sentiment analysis in twitter](https://md.research.elamur.ai/free/en/papers/rosenthal-2014-semeval-2014-task-9-sentiment/). In SemEval@COLING, 2014.
4. Bo Pang and Lillian Lee. Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2007. [↗](https://doi.org/10.1561/9781601981516)
5. Luís Sarmento, Paula Carvalho, Mário J Silva, and Eugénio De Oliveira. Automatic creation of a reference corpus for political opinion mining in user-generated content. In Proceedings of the 1st international CIKM workshop on Topic-sentiment analysis for mass opinion, 2009. [find](https://scholar.google.com/scholar?q=Lu%C3%ADs+Sarmento%2C+Paula+Carvalho%2C+M%C3%A1rio+J+Silva%2C+and+Eug%C3%A9nio+De+Oliveira%2E+Automatic+creation+of+a+reference+corpus+for+political+opinion+mining+in+user%2Dgenerated+content%2E+In+Proceedings+of+the+1st+intern)
6. Cynthia Van Hee, Els Lefever, and Véronique Hoste. [Semeval-2018 task 3: Irony detection in english tweets](https://md.research.elamur.ai/free/en/papers/hee-2018-semeval-2018-task-3-irony/). In SemEval@NAACL-HLT, 2018.
7. Tianxiao Shen, Tao Lei, Regina Barzilay, and Tommi S. Jaakkola. [Style transfer from non-parallel text by cross-alignment](https://md.research.elamur.ai/free/en/papers/shen-2017-style-transfer-non-parallel-text/). In NIPS, 2017. [↗](https://doi.org/10.48550/arxiv.1705.09655)
8. Shrimai Prabhumoye, Yulia Tsvetkov, Ruslan Salakhutdinov, and Alan W. Black. [Style transfer through back-translation](https://md.research.elamur.ai/free/en/papers/prabhumoye-2018-style-transfer-through-back-translation/). In ACL (1), 2018.
9. Fuli Luo, Peng Li, Jie Zhou, Pengcheng Yang, Baobao Chang, Xu Sun, and Zhifang Sui. [A dual reinforcement learning framework for unsupervised text style transfer](https://md.research.elamur.ai/free/en/papers/luo-2019-dual-reinforcement-learning-framework-unsupervised/). In IJCAI, 2019.
10. Juncen Li, Robin Jia, He He, and Percy Liang. [Delete, retrieve, generate: a simple approach to sentiment and style transfer](https://md.research.elamur.ai/free/en/papers/li-2018-delete-retrieve-generate-simple-approach/). In NAACL-HLT, 2018. See the examples in Table 4. [↗](https://doi.org/10.48550/arxiv.1804.06437)
11. Jingjing Xu, Xu Sun, Qi Zeng, Xiaodong Zhang, Xuancheng Ren, Houfeng Wang, and Wenjie Li. [Unpaired sentiment-to-sentiment translation: A cycled reinforcement learning approach](https://md.research.elamur.ai/free/en/papers/xu-2018-unpaired-sentiment-sentiment-translation-cycled/). In ACL (1), 2018. [↗](https://doi.org/10.60692/kzdbq-6en85)
12. Guillaume Lample, Sandeep Subramanian, Eric Michael Smith, Ludovic Denoyer, Marc’Aurelio Ranzato, and Y-Lan Boureau. Multiple-attribute text rewriting. In ICLR (Poster), 2019. [find](https://scholar.google.com/scholar?q=Guillaume+Lample%2C+Sandeep+Subramanian%2C+Eric+Michael+Smith%2C+Ludovic+Denoyer%2C+Marc%E2%80%99Aurelio+Ranzato%2C+and+Y%2DLan+Boureau%2E+Multiple%2Dattribute+text+rewriting%2E+In+ICLR+%28Poster%29%2C+2019%2E)
13. Zhenxin Fu, Xiaoye Tan, Nanyun Peng, Dongyan Zhao, and Rui Yan. [Style transfer in text: Exploration and evaluation](https://md.research.elamur.ai/free/en/papers/fu-2017-style-transfer-text-exploration-evaluation/). In AAAI, 2018.
14. Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P. Xing. [Toward controlled generation of text](https://md.research.elamur.ai/free/en/papers/hu-2017-toward-controlled-generation-text/). In ICML, 2017. [↗](https://doi.org/10.48550/arxiv.1703.00955)
15. Omid Rohanian, Shiva Taslimipoor, Richard Evans, and Ruslan Mitkov. [WLV at semeval-2018 task 3: Dissecting tweets in search of irony](https://md.research.elamur.ai/free/en/papers/rohanian-2018-wlv-semeval-2018-task-3/). In SemEval@NAACL-HLT, 2018.
16. Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 1997. [↗](https://doi.org/10.1162/neco.1997.9.8.1735)
17. Aniruddha Ghosh and Tony Veale. [Ironymagnet at semeval-2018 task 3: A siamese network for irony detection in social media](https://md.research.elamur.ai/free/en/papers/ghosh-2018-ironymagnet-semeval-2018-task-3/). In SemEval@NAACL-HLT, 2018.
18. Christos Baziotis, Athanasiou Nikolaos, Alexandra Chronopoulou, Athanasia Kolovou, Georgios Paraskevopoulos, Nikolaos Ellinas, Shrikanth Narayanan, and Alexandros Potamianos. NTUA-SLP at semeval-2018 task 1: Predicting affective content in tweets with deep attentive rnns and transfer learning. In SemEval@NAACL-HLT, 2018. [find](https://scholar.google.com/scholar?q=Christos+Baziotis%2C+Athanasiou+Nikolaos%2C+Alexandra+Chronopoulou%2C+Athanasia+Kolovou%2C+Georgios+Paraskevopoulos%2C+Nikolaos+Ellinas%2C+Shrikanth+Narayanan%2C+and+Alexandros+Potamianos%2E+NTUA%2DSLP+at+semeval%2D2018+)
19. Chuhan Wu, Fangzhao Wu, Sixing Wu, Junxin Liu, Zhigang Yuan, and Yongfeng Huang. Thu_ngn at semeval-2018 task 3: Tweet irony detection with densely connected LSTM and multi-task learning. In SemEval@NAACL-HLT, 2018. [find](https://scholar.google.com/scholar?q=Chuhan+Wu%2C+Fangzhao+Wu%2C+Sixing+Wu%2C+Junxin+Liu%2C+Zhigang+Yuan%2C+and+Yongfeng+Huang%2E+Thu%5Fngn+at+semeval%2D2018+task+3%3A+Tweet+irony+detection+with+densely+connected+LSTM+and+multi%2Dtask+learning%2E+In+SemEval%40N)
20. Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data. In ACL (1), 2016. [find](https://scholar.google.com/scholar?q=Rico+Sennrich%2C+Barry+Haddow%2C+and+Alexandra+Birch%2E+Improving+neural+machine+translation+models+with+monolingual+data%2E+In+ACL+%281%29%2C+2016%2E)
21. Yoon Kim. [Convolutional neural networks for sentence classification](https://md.research.elamur.ai/free/en/papers/kim-2014-convolutional-neural-networks-sentence-classification/). In EMNLP, 2014.
22. Ronald J. Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 1992. [↗](https://doi.org/10.1007/bf00992696)
23. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. [Attention is all you need](https://md.research.elamur.ai/free/en/papers/vaswani-2017-attention-all-need/). In NIPS, 2017.
24. Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. [↗](https://doi.org/10.48550/arxiv.1412.6980)
25. Lajanugen Logeswaran, Honglak Lee, and Samy Bengio. Content preserving text generation with attribute controls. In NeurIPS, 2018. [↗](https://doi.org/10.48550/arxiv.1811.01135)
26. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. [Bleu: a method for automatic evaluation of machine translation](https://md.research.elamur.ai/free/en/papers/papineni-2002-bleu-method-automatic-evaluation-machine/). In ACL, 2002.

*Updated 24 October 2026*