Elamur Research⌕
☰
home / papers / 2024 / How Much Do Robots Understand Rudeness? Challenges in Human-Robot Interaction

How Much Do Robots Understand Rudeness? Challenges in Human-Robot Interaction

Paper · 2024 · Michael Andrew Orme, Yanchao Yu and Zhiyuan Tan

Summary

Robots and assistants are moving into homes and classrooms, where they talk with children, so they must avoid producing or repeating offensive speech. This paper asks how well common cleaning and detection tools actually handle rudeness. The authors use the Cornell Movie Dialog Corpus as a proxy for realistic human-human conversation. They flag profane words with the Better-profanity word list and build three cleaned versions of the corpus:

  • deleting the flagged words,
  • paraphrasing only those words,
  • rewriting whole sentences with a T5-based paraphraser.

A DialoGPT-small dialogue model is fine-tuned on each version. The versions are compared with the original on two things: how much rudeness remains, and how close the responses stay in meaning to the originals. Large language models, hate speech detectors and sentiment analysers serve as judges.

The overall finding is that these quick fixes are weaker than they look. Rudeness often hides inside idioms and euphemisms that word lists miss. Whether something is offensive also depends on context and on the background of the people involved. This matters for anyone building conversational agents for vulnerable audiences.

Key points

  • The source data is the Cornell Movie Dialog Corpus (220,579 exchanges between character pairs). Its estimated share of inappropriate content depends heavily on the detector: about 9% by profanity detection, 12.08% by hate speech detection and 63.06% by sentiment analysis.
  • Profanity is detected with Better-profanity, a word-by-word lookup against a human-compiled list. The authors note it is subjective and blind to context.
  • Word and sentence paraphrasing use humarin's chatgpt_paraphraser_on_T5_base. The authors also fine-tuned T5-Large (t5-v1_1-large) on the same chatgpt-paraphrases dataset. It brought no significant gain, sometimes left more rudeness than the unparaphrased text, and was judged not cost-effective.
  • Dialogue models are DialoGPT-small, fine-tuned with the HuggingFace Trainer: batch size 4, Adam, learning rate 1e-4, 5 epochs, cross-entropy loss.
  • The two central obstacles identified are offensive meaning concealed in idiomatic or euphemistic phrasing, and offensiveness that varies with context, gender, ethnicity and personal background.

Interesting ideas

  • I1. Estimating the same property with several detectors of different kinds exposes how unstable the estimate is (9% vs 63% on one corpus), so a single judge's number should not be trusted.
  • I2. A cleaning step is judged on two axes at once: how much of the unwanted property survives, and how much of the original meaning is kept.
  • I3. Re-scoring a rewriter's output with the same detector revealed that the larger paraphraser sometimes added rudeness instead of removing it.
  • I4. The three cleaning methods form a ladder of intervention: delete the word, replace the word, or rewrite the sentence.
  • I5. Flat word lists fail because offence lives in idioms and in who says what to whom, not in individual tokens.

For Mova

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

  • I1 (learn): Treat rudeness or sentiment judgements like silver parses, accepting a label only when independent witnesses agree and reporting their disagreement rate as a negative-control signal.
  • I2 (back): The text → graph → text compressor and any rewriting could be scored on meaning preservation over the frame graph alongside a target-property check, rather than on surface fluency alone.
  • I3 (back): Every generation pass could be re-read by Mova's own pipeline to verify it did not introduce what it was meant to eliminate.
  • I4 (back): Generation could offer graded edits on specific tree nodes versus whole-clause regeneration from frames, choosing the smallest edit that fixes the problem.
  • I5 (frames): The prag crate could flag insults through speech-act and implicature analysis of speaker, addressee and idiomatic constructions, giving an explainable rudeness decision instead of a lexicon hit.

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

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksCommunication · 97 works

Cite

Cite as Michael Andrew Orme, Yanchao Yu and Zhiyuan Tan. 2024. How Much Do Robots Understand Rudeness? Challenges in Human-Robot Interaction. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024).

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

@inproceedings{orme2024how,
  title = {{How Much Do Robots Understand Rudeness? Challenges in Human-Robot Interaction}},
  author = {Michael Andrew Orme, Yanchao Yu and Zhiyuan Tan},
  booktitle = {Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  year = {2024},
  doi = {10.63317/3x7a4xg5r77f},
  url = {https://doi.org/10.63317/3x7a4xg5r77f}
}

References

31 references, 9 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Sai Saket Aluru, Binny Mathew, Punyajoy Saha, and Animesh Mukherjee. 2020a. Deep learning models for multilingual hate speech detection. arXiv preprint arXiv:2004.06465. ↗
  2. Sai Saketh Aluru, Binny Mathew, Punyajoy Saha, and Animesh Mukherjee. 2020b. A deep dive into multilingual hate speech classification. In Machine Learning and Knowledge Discovery in Databases. Applied Data Science and Demo Track: European Conference, ECML PKDD 2020, Ghent, Belgium, September 14–18, 2020, Proceedings, Part V, page 423–439, Berlin, Heidelberg. Springer-Verlag. find
  3. Dimosthenis Antypas and Jose Camacho-Collados. 2023a. Robust hate speech detection in social media: A cross-dataset empirical evaluation. In The 7th Workshop on Online Abuse and Harms (WOAH), pages 231–242, Toronto, Canada. Association for Computational Linguistics. find
  4. Dimosthenis Antypas and Jose Camacho-Collados. 2023b. Robust hate speech detection in social media: A cross-dataset empirical evaluation. In The 7th Workshop on Online Abuse and Harms (WOAH), pages 231–242, Toronto, Canada. Association for Computational Linguistics. find
  5. Apoorva Chaudhari, Palak Davda, Monil Dand, and Surekha Dholay. 2021. Profanity detection and removal in videos using machine learning. In 2021 6th International Conference on Inventive Computation Technologies (ICICT), pages 572– 576. find
  6. Jithin Cheriyan, Bastin Tony Roy Savarimuthu, and Stephen Cranefield. 2021. Towards offensive language detection and reduction in four software engineering communities. In Proceedings of the 25th International Conference on Evaluation and Assessment in Software Engineering, EASE ’21, page 254–259, New York, NY, USA. Association for Computing Machinery. find
  7. Sarah M Coyne, Laura A Stockdale, David A Nelson, and Ashley Fraser. 2011. Profanity in media associated with attitudes and behavior regarding profanity use and aggression. Pediatrics, 128(5):867–872. ↗
  8. David Dale, Anton Voronov, Daryna Dementieva, Varvara Logacheva, Olga Kozlova, Nikita Semenov, and Alexander Panchenko. 2021. Text detoxification using large pre-trained neural models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7979–7996, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics. find
  9. Cristian Danescu-Niculescu-Mizil and Lillian Lee. 2011. Chameleons in imagined conversations: A new approach to understanding coordination of linguistic style in dialogs. Proceedings of the ACL workshop on Cognitive Modeling and Computational Linguistics, pages 76–87. ↗
  10. Mithun Das, Somnath Banerjee, and Animesh Mukherjee. 2022a. Data bootstrapping approaches to improve low resource abusive language detection for indic languages. In Proceedings of the 33rd ACM Conference on Hypertext and Social Media, HT ’22, page 32–42, New York, NY, USA. Association for Computing Machinery. find
  11. Mithun Das, Somnath Banerjee, and Animesh Mukherjee. 2022b. Data bootstrapping approaches to improve low resource abusive language detection for indic languages. arXiv preprint arXiv:2204.12543. ↗
  12. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. ↗
  13. Vanessa Hahn, Dana Ruiter, Thomas Kleinbauer, and Dietrich Klakow. 2021. Modeling profanity and hate speech in social media with semantic subspaces. In Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 6–16, Online. Association for Computational Linguistics. find
  14. Jochen Hartmann, Mark Heitmann, Christian Siebert, and Christina Schamp. 2023. More than a feeling: Accuracy and application of sentiment analysis. International Journal of Research in Marketing, 40(1):75–87. ↗
  15. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692. ↗
  16. Daniel Loureiro, Francesco Barbieri, Leonardo Neves, Luis Espinosa Anke, and Jose Camacho-collados. 2022. TimeLMs: Diachronic language models from Twitter. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 251–260, Dublin, Ireland. Association for Computational Linguistics. ↗
  17. Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2020. Hatexplain: A benchmark dataset for explainable hate speech detection. arXiv preprint arXiv:2012.10289. ↗
  18. Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2021. Hatexplain: A benchmark dataset for explainable hate speech detection. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 14867– 14875. find
  19. George A. Miller and Walter G. Charles. 1991. Contextual correlates of semantic similarity. Language and Cognitive Processes, 6(1):1–28. ↗
  20. Cicero Nogueira dos Santos, Igor Melnyk, and Inkit Padhi. 2018. Fighting offensive language on social media with unsupervised text style transfer. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 189–194, Melbourne, Australia. Association for Computational Linguistics. ↗
  21. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. The Journal of Machine Learning Research, 21(1):5485–5551. ↗
  22. Catherine Rondina. 2005. Rudeness: Deal with it if You Please. James Lorimer & Company. find
  23. M. H. J. Smakman, E. A. Konijn, and P. A. Vogt. 2022. Do robotic tutors compromise the social-emotional development of children?. Frontiers in Robotics and AI, 9. ↗
  24. Hui-Po Su, Zhen-Jie Huang, Hao-Tsung Chang, and Chuan-Jie Lin. 2017. Rephrasing profanity in Chinese text. In Proceedings of the First Workshop on Abusive Language Online, pages 18–24, Vancouver, BC, Canada. Association for Computational Linguistics. find
  25. Minh Tran, Yipeng Zhang, and Mohammad Soleymani. 2020. Towards a friendly online community: An unsupervised style transfer framework for profanity redaction. In Proceedings of the 28th International Conference on Computational Linguistics, pages 2107–2114, Barcelona, Spain (Online). International Committee on Computational Linguistics. find
  26. Bertie Vidgen, Tristan Thrush, Zeerak Waseem, and Douwe Kiela. 2021a. Learning from the worst: Dynamically generated datasets to improve online hate detection. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1667–1682, Online. Association for Computational Linguistics.
  27. Bertie Vidgen, Tristan Thrush, Zeerak Waseem, and Douwe Kiela. 2021b. Learning from the worst: Dynamically generated datasets to improve online hate detection. In ACL.
  28. Dirk von Grünigen, Ralf Grubenmann, Fernando Benites, Pius von Däniken, and Mark Cieliebak. 2018. spmmmp at germeval 2018 shared task: Classification of offensive content in tweets using convolutional neural networks and gated recurrent units. find
  29. Michael Wiegand and Melanie Siegel. 2018. Overview of the germeval 2018 shared task on the identification of offensive language. find
  30. Zhi Xu and Sencun Zhu. 2010. Filtering offensive language in online communities using grammatical relations. 7th Annual Collaboration, Electronic Messaging, Anti-Abuse and Spam Conference, CEAS 2010. find
  31. Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020. DIALOGPT : Large-scale generative pre-training for conversational response generation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 270–278, Online. Association for Computational Linguistics. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/orme-2024-how-much-do-robots-understand/