---
title: "Polite Task-oriented Dialog Agents: To Generate or to Rewrite?"
description: "For it they build Domain-Extended (DE) Word2Vec embeddings trained on MMD plus the Stanford… 21 references, cited by 4 papers here."
authors: ["Diogo Silva", "David Semedo", "João Magalhães"]
year: 2022
venue: "Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis"
source: "https://aclanthology.org/2022.wassa-1.34/"
doi: "10.18653/v1/2022.wassa-1.34"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/silva-2022-polite-task-oriented-dialog-agents/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/silva-2022-polite-task-oriented-dialog-agents/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2022](https://md.research.elamur.ai/free/en/papers/2022/page/5/) / Polite Task-oriented Dialog Agents: To Generate or to Rewrite?

# Polite Task-oriented Dialog Agents: To Generate or to Rewrite?

**Authors:** Diogo Silva, [David Semedo](https://md.research.elamur.ai/free/en/authors/david-semedo/) and [João Magalhães](https://md.research.elamur.ai/free/en/authors/joao-magalhaes/) · **Year:** 2022 · **Venue:** [Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis](https://md.research.elamur.ai/free/en/venues/wassa/) ([Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis 2022](https://md.research.elamur.ai/free/en/venues/wassa/2022/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2022.wassa-1.34/) · [PDF](https://aclanthology.org/2022.wassa-1.34.pdf) · [DOI 10.18653/v1/2022.wassa-1.34](https://doi.org/10.18653/v1/2022.wassa-1.34) · [OpenAlex](https://openalex.org/W4285158411)

**Citations:** 5 citations (OpenAlex, 2026-10-14) · [4](https://md.research.elamur.ai/free/en/papers/silva-2022-polite-task-oriented-dialog-agents/cited-by/) in our library, of the 141,840 papers whose bibliographies we read

## Summary
The paper asks how to make task-oriented dialog agents sound polite in a narrow commercial domain, here fashion shopping, without damaging the factual content of their replies. Existing politeness methods were built on general-purpose data and lack labels for specialised vocabulary. The authors compare two families of methods under domain transfer. The first is generative: Polite-RL, a Seq2Seq model rewarded by a politeness classifier. For it they build Domain-Extended (DE) Word2Vec embeddings trained on MMD plus the Stanford Politeness Corpus. The second is rewriting: Tag-and-Generate, which tags spots for style markers and fills them in. They test four ways to choose and train its markers, including filtering out product nouns and reusing markers learned on Enron. Automatic metrics and a user study both favour rewriting. It balances added politeness against content preservation better, because generation must solve answer production and tone at the same time. This matters for any agent whose tone reflects a brand while its answers must stay correct.

## Key points
- The widely used politeness classifier from Danescu-Niculescu-Mizil et al. (2013) lacks coverage for roughly 15k terms that appear in the fashion dialog corpus MMD (Saha et al., 2018).
- DE embeddings come from a Word2Vec model trained on MMD concatenated with the original politeness corpus. They replace the Google News vectors used in the original Polite-RL.
- The four rewriting setups are:
  - **RW-Enron:** trained on Enron.
  - **RW-Fashion:** fashion data labelled by classifier score, with polite above 0.9 and rude between 0.5 and 0.6.
  - **RW-Fashion-Clean:** the same, but product nouns such as "scarf" are blocked as markers.
  - **RW-Mixed:** markers learned on Enron, training done on fashion data.
- Models gained most on short rude utterances and changed neutral ones very little. The strategies they used most were Gratitude and Positive Lexicon.
- On the Enron test, higher classifier scores from models trained on fashion data were misleading. The outputs were often just pasted fashion snippets, so the politeness score rose without any real improvement in tone.

## Interesting ideas
- Keep tone adjustment as a separate rewriting pass instead of mixing it into answer generation, so content correctness and style can each be controlled and debugged alone.
- Before mining style markers by TF-IDF contrast, filter out domain content words so topic vocabulary is not mistaken for a style signal.
- Learn style markers on a richly labelled source domain, then train the insertion model on target-domain text (the RW-Mixed idea), so noisy target labels do not pollute the marker set.
- A rising classifier score does not prove better style; reading the outputs showed that gains came from pasted domain snippets.
- A marker-based rewriter can only change text that already contains markers, so neutral input stays almost untouched.

## For Mova

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

- **I1** (back): After generating the core answer from the meaning graph, a separate rewriting step could add politeness or register while the graph-derived content stays fixed and checkable.
- **I2** (learn): When inducing repair rules or style cues from contrasting corpora, a stoplist of entity and product nouns from the world layer can stop topic leakage into the learned rules.
- **I3** (learn): Speech-act or register cues could be harvested from a well-annotated corpus and then applied to new genres such as fables or school texts without relabelling.
- **I4** (learn): Evaluation of any style or pragmatic component should include manual inspection or negative controls, so metric gaming by irrelevant insertions is caught.
- **I5** (frames): A politeness or speech-act detector in prag should treat the absence of markers as its own case, not as a low score on a polite–rude scale.

*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

- [Universidade Nova de Lisboa](https://md.research.elamur.ai/free/en/universities/universidade-nova-de-lisboa/) — University, Lisbon, Portugal · 3 works · 1,891 citations

Also on the byline: [Laboratório para a Ciência da Computação e Informática](https://ror.org/058e2tg06).

📚 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): [🗨️ Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) · 401 works

## Cite

Cite as: Diogo Silva, David Semedo and João Magalhães. 2022. Polite Task-oriented Dialog Agents: To Generate or to Rewrite?. Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis.

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

```bibtex
@inproceedings{silva2022polite,
  title = {{Polite Task-oriented Dialog Agents: To Generate or to Rewrite?}},
  author = {Diogo Silva, David Semedo and João Magalhães},
  booktitle = {Proceedings of the 12th Workshop on Computational Approaches to Subjectivity, Sentiment \& Social Media Analysis},
  year = {2022},
  doi = {10.18653/v1/2022.wassa-1.34},
  url = {https://doi.org/10.18653/v1/2022.wassa-1.34}
}
```

- [← previous Nucleus Composition in Transition-based Dependency Parsing](https://md.research.elamur.ai/free/en/papers/nivre-2022-nucleus-composition-transition-based-dependency/)
- [next → Restoring Hebrew Diacritics Without a Dictionary](https://md.research.elamur.ai/free/en/papers/gershuni-2022-restoring-hebrew-diacritics-without-dictionary/)
- [↑ back to papers of 2022](https://md.research.elamur.ai/free/en/papers/2022/page/5/)

## References

21 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Ardith J. Meier. 1995. Defining politeness: Universality in appropriateness. Language Sciences, 17:345–356. [↗](https://doi.org/10.1016/0388-0001(95)00019-4)
2. Malika Aubakirova and Mohit Bansal. 2016. [Interpreting neural networks to improve politeness comprehension](https://md.research.elamur.ai/free/en/papers/aubakirova-2016-interpreting-neural-networks-improve-politeness/). In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2035–2041, Austin, Texas. Association for Computational Linguistics.
3. Tomas Mikolov, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. 2013. [Efficient estimation of word representations in vector space](https://md.research.elamur.ai/free/en/papers/mikolov-2013-efficient-estimation-word-representations-vector/). In ICLR. [↗](https://doi.org/10.48550/arxiv.1301.3781)
4. Tong Niu and Mohit Bansal. 2018. Polite dialogue generation without parallel data. Transactions of the Association for Computational Linguistics, 6:373–389. [↗](https://doi.org/10.1162/tacl_a_00027)
5. Francesca Bargiela-Chiappini. 2003. Face and politeness: new (insights) for old (concepts). Journal of Pragmatics, 35(10):1453–1469. [↗](https://doi.org/10.1016/s0378-2166(02)00173-x)
6. Penelope Brown, Stephen Levinson, and E Goody. 1978. Universal in Language Usage: Politeness Phenomena, volume 8, pages 56–311. [find](https://scholar.google.com/scholar?q=Penelope+Brown%2C+Stephen+Levinson%2C+and+E+Goody%2E+1978%2E+Universal+in+Language+Usage%3A+Politeness+Phenomena%2C+volume+8%2C+pages+56%E2%80%93311%2E)
7. Penelope Brown, Stephen C Levinson, and Stephen C Levinson. 1987. Politeness: Some universals in language usage, volume 4. Cambridge university press. [find](https://scholar.google.com/scholar?q=Penelope+Brown%2C+Stephen+C+Levinson%2C+and+Stephen+C+Levinson%2E+1987%2E+Politeness%3A+Some+universals+in+language+usage%2C+volume+4%2E+Cambridge+university+press%2E)
8. Kyunghyun Cho, Bart van Merriënboer, Dzmitry Bahdanau, and Yoshua Bengio. 2014. [On the properties of neural machine translation: Encoder–decoder approaches](https://md.research.elamur.ai/free/en/papers/cho-2014-properties-neural-machine-translation-encoder/). In Proceedings of SSST-8, Eighth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 103–111, Doha, Qatar. Association for Computational Linguistics.
9. Cristian Danescu-Niculescu-Mizil, Moritz Sudhof, Dan Jurafsky, Jure Leskovec, and Christopher Potts. 2013. [A computational approach to politeness with application to social factors](https://md.research.elamur.ai/free/en/papers/mizil-2013-computational-approach-politeness-application-social/). In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 250–259, Sofia, Bulgaria. Association for Computational Linguistics. [↗](https://doi.org/10.48550/arxiv.1306.6078)
10. Michael J. Denkowski and Alon Lavie. 2011. [Meteor 1.3: Automatic metric for reliable optimization and evaluation of machine translation systems](https://md.research.elamur.ai/free/en/papers/denkowski-2011-meteor-1-3-automatic-metric/). In WMT@EMNLP.
11. Mauajama Firdaus, Asif Ekbal, and Pushpak Bhattacharyya. 2020. Incorporating politeness across languages in customer care responses: Towards building a multi-lingual empathetic dialogue agent. In Proceedings of the 12th Language Resources and Evaluation Conference, pages 4172– 4182, Marseille, France. European Language Resources Association. [find](https://scholar.google.com/scholar?q=Mauajama+Firdaus%2C+Asif+Ekbal%2C+and+Pushpak+Bhattacharyya%2E+2020%2E+Incorporating+politeness+across+languages+in+customer+care+responses%3A+Towards+building+a+multi%2Dlingual+empathetic+dialogue+agent%2E+In+Proc)
12. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. [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.
13. Amrita Saha, Mitesh M. Khapra, and Karthik Sankaranarayanan. 2018. Towards building large scale multimodal domain-aware conversation systems. In Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium on Educational Advances in Artificial Intelligence (EAAI-18), New Orleans, Louisiana, USA, February 2-7, 2018, pages 696–704. AAAI Press. [↗](https://doi.org/10.1609/aaai.v32i1.11331)
14. Ricardo Gamelas Sousa, Pedro Miguel Ferreira, Pedro Moreira Costa, Pedro Azevedo, Joao Paulo Costeira, Carlos Santiago, Joao Magalhaes, David Semedo, Rafael Ferreira, Alexander I. Rudnicky, and Alexander Georg Hauptmann. 2021. IFetch: Multimodal Conversational Agents for the Online Fashion Marketplace, page 25–26. Association for Computing Machinery, New York, NY, USA. [find](https://scholar.google.com/scholar?q=Ricardo+Gamelas+Sousa%2C+Pedro+Miguel+Ferreira%2C+Pedro+Moreira+Costa%2C+Pedro+Azevedo%2C+Joao+Paulo+Costeira%2C+Carlos+Santiago%2C+Joao+Magalhaes%2C+David+Semedo%2C+Rafael+Ferreira%2C+Alexander+I%2E+Rudnicky%2C+and+Alexan)
15. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. [Attention is all you need](https://md.research.elamur.ai/free/en/papers/vaswani-2017-attention-all-need/). In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, page 6000–6010, Red Hook, NY, USA. Curran Associates Inc.
16. Richard J. Watts. 2019. 2. Linguistic politeness and politic verbal behaviour: Reconsidering claims for universality, pages 43–70. De Gruyter Mouton. [find](https://scholar.google.com/scholar?q=Richard+J%2E+Watts%2E+2019%2E+2%2E+Linguistic+politeness+and+politic+verbal+behaviour%3A+Reconsidering+claims+for+universality%2C+pages+43%E2%80%9370%2E+De+Gruyter+Mouton%2E)
17. Liye Fu, Susan Fussell, and Cristian Danescu-Niculescu-Mizil. 2020. [Facilitating the communication of politeness through fine-grained paraphrasing](https://md.research.elamur.ai/free/en/papers/fu-2020-facilitating-communication-politeness-through-fine/). In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5127–5140, Online. Association for Computational Linguistics.
18. Hitesh Golchha, M. Firdaus, Asif Ekbal, and P. Bhattacharyya. 2019. Courteously yours: Inducing courteous behavior in customer care responses using reinforced pointer generator network. In NAACL-HLT. [find](https://scholar.google.com/scholar?q=Hitesh+Golchha%2C+M%2E+Firdaus%2C+Asif+Ekbal%2C+and+P%2E+Bhattacharyya%2E+2019%2E+Courteously+yours%3A+Inducing+courteous+behavior+in+customer+care+responses+using+reinforced+pointer+generator+network%2E+In+NAACL%2DHLT%2E)
19. Bryan Klimt and Yiming Yang. 2004. Introducing the enron corpus. In CEAS. [find](https://scholar.google.com/scholar?q=Bryan+Klimt+and+Yiming+Yang%2E+2004%2E+Introducing+the+enron+corpus%2E+In+CEAS%2E)
20. Chin-Yew Lin. 2004. [Rouge: A package for automatic evaluation of summaries](https://md.research.elamur.ai/free/en/papers/lin-2004-rouge-package-automatic-evaluation-summaries/). In ACL 2004.
21. Aman Madaan, Amrith Setlur, Tanmay Parekh, Barnabas Poczos, Graham Neubig, Yiming Yang, Ruslan Salakhutdinov, Alan W Black, and Shrimai Prabhumoye. 2020. [Politeness transfer: A tag and generate approach](https://md.research.elamur.ai/free/en/papers/madaan-2020-politeness-transfer-tag-generate-approach/). In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics. Politeness Strategies Using Honorifics: How can I help you Ms Smith? Description: Through the usage of Honorifics, the speaker conveys respect towards another person. Gratitude: Thank you very much for your help. Description: One of the easiest ways of sounding polite is through the expression of gratitude. Tag Questions: Take a seat, won’t you? Description: Tag Questions at the end of a utterance can define the tone used, making it sound more friendly or the opposite, depending on the usage. BLEU Politeness Reference(96, 2.0) Model 1(32, 2.0) Model 2 (128, 2.0) Model 3 (96, 5.0) Model 4 (96, 10.0) 66.30 49.64 63.60 33.56 29.41 64.47 72.40 60.24 75.21 78.77

*Updated 24 October 2026*