---
title: "Enhancing LLM-based Arabic Negotiation by Fine Tuning on"
description: "The paper looks at why general-purpose LLMs handle Arabic bargaining dialogues poorly. 19 references."
authors: ["Yasmine A Abu Adla", "Hazem Hajj", "Shady Elbassuoni", "Khaled Shaban", "Wassim El Hajj"]
year: 2024
venue: "Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024)"
source: "https://aclanthology.org/2024.icnlsp-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/adla-2024-enhancing-llm-based-arabic-negotiation/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/adla-2024-enhancing-llm-based-arabic-negotiation/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2024](https://md.research.elamur.ai/free/en/papers/2024/page/16/) / Enhancing LLM-based Arabic Negotiation by Fine Tuning on Dialogue Shortcomings

# Enhancing LLM-based Arabic Negotiation by Fine Tuning on Dialogue Shortcomings

**Authors:** [Yasmine A Abu Adla](https://md.research.elamur.ai/free/en/authors/yasmine-a-abu-adla/), [Hazem Hajj](https://md.research.elamur.ai/free/en/authors/hazem-hajj/), [Shady Elbassuoni](https://md.research.elamur.ai/free/en/authors/shady-elbassuoni/), [Khaled Shaban](https://md.research.elamur.ai/free/en/authors/khaled-shaban/) and [Wassim El Hajj](https://md.research.elamur.ai/free/en/authors/wassim-el-hajj/) · **Year:** 2024 · **Venue:** [Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024)](https://md.research.elamur.ai/free/en/venues/icnlsp/) ([International Conference on Natural Language and Speech Processing 2024](https://md.research.elamur.ai/free/en/venues/icnlsp/2024/)) · **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/2024.icnlsp-1.34/) · [PDF](https://aclanthology.org/2024.icnlsp-1.34.pdf)

## Summary
The paper looks at why general-purpose LLMs handle Arabic bargaining dialogues poorly. The authors first probed several models (Gemini, LLaMA 2, BARD, GPT-3.5 Turbo) with a role-play item-splitting game. They found three recurring weaknesses:
- miscalculated values,
- broken chain-of-thought (COT) reasoning,
- unstable pronoun reference in Arabic.

To repair these, they had GPT-4 generate a synthetic corpus of 1,400 Arabic negotiation scenarios. The scenarios were written to reflect features of Arabic bargaining such as indirectness, rich morphology and complex pronouns. GPT-3.5 Turbo was then fine-tuned on this corpus.

The evaluation compared the fine-tuned model with the base model and with GPT-4 reference behaviour on three measures: dialogue length, strategic depth and arithmetic accuracy. The tuned model produced longer exchanges and adjusted offers more strategically. It also stayed inside the given scenario, for example keeping the specified currency.

The work matters because negotiation agents have been studied almost only in English. It shows a cheap path for an underrepresented language: diagnose the failures, then synthesize targeted data. The dataset is announced for public release.

## Key points
- Diagnostic setup: two characters, Amani and Samir, split a book, three hats and a ball under different private valuations. The game ends with an explicit agreement and is scored, and it was run ten times with GPT-3.5 Turbo playing both sides.
- Failure categories identified across models:
  - arithmetic and logical errors,
  - lapses in common-sense reasoning,
  - inconsistent pronoun use in Arabic.
- Training resource: 1,400 GPT-4-synthesized Arabic negotiation dialogues, each produced from scenario-specific prompts. GPT-3.5 Turbo was fine-tuned on them.
- Evaluation axes:
  - dialogue length,
  - strategic depth,
  - arithmetic accuracy.

  The fine-tuned model is reported to beat the baseline on engagement and strategy, and to exceed GPT-4 on some aspects.
- Qualitative contrast: the base model drifted between topics and once switched from Saudi Riyal to dollars in a tailored-suit deal. The fine-tuned model kept to the scenario.

## Interesting ideas
- Build the training data from a diagnosed failure list instead of generically: first catalogue exact error types (sums, reasoning chains, pronouns), then generate examples aimed at each one, so effort goes where the model actually breaks.
- A closed-world bargaining game with fixed private valuations and a final score gives a checkable ground truth for whether the dialogue's arithmetic is right.
- Tracking each party's inventory turn by turn turns a conversation into a sequence of state changes that can be audited.
- Use scenario-constraint adherence as a measurable signal, for example whether the stated currency or topic is kept, since such drift is easy to detect and reveals lost context.
- Pronoun consistency across turns is treated as a distinct failure mode of dialogue, separate from reasoning errors.

## For Mova

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

- **I1** (learn): LLM-teacher tiers could be asked to produce silver examples only for error classes found by Mova's own evaluation, instead of broad random text.
- **I2** (answer): similar toy exchange worlds (who holds which items at what value) could serve as maths-word-problem tests, where the final holdings are computed and verified exactly.
- **I3** (world): holdings and event order in the world model could be checked against stated totals after every transfer, flagging texts or readings where quantities fail to add up.
- **I4** (back): generated text from the meaning graph could be checked automatically for preserved units, currencies and entities as a negative control on the text → graph → text compressor.
- **I5** (frames): coreference in multi-speaker dialogue (I/you swapping between turns) deserves its own test set and score, kept apart from narrative coreference.

*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 (2): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works

🧭 Topics (2): [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 3,066 works · [🗨️ Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) · 654 works

## Cite

Cite as: Yasmine A Abu Adla, Hazem Hajj, Shady Elbassuoni, Khaled Shaban and Wassim El Hajj. 2024. Enhancing LLM-based Arabic Negotiation by Fine Tuning on Dialogue Shortcomings. Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024).

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

```bibtex
@inproceedings{adla2024enhancing,
  title = {{Enhancing LLM-based Arabic Negotiation by Fine Tuning on Dialogue Shortcomings}},
  author = {Yasmine A Abu Adla, Hazem Hajj, Shady Elbassuoni, Khaled Shaban and Wassim {El Hajj}},
  booktitle = {Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024)},
  year = {2024},
  url = {https://aclanthology.org/2024.icnlsp-1.34/}
}
```

- [← previous English Verbs of Directional Vision:Compatibility with Designations of a Qualitative Attribute](https://md.research.elamur.ai/free/en/papers/antonova-2024-english-verbs-directional-vision-compatibility/)
- [next → Enhancing Medical Support in the Arabic Language Through Personalized ChatGPT Assistance](https://md.research.elamur.ai/free/en/papers/issa-2024-enhancing-medical-support-arabic-language/)
- [↑ back to papers of 2024](https://md.research.elamur.ai/free/en/papers/2024/page/16/)

## References

19 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Sahar Abdelnabi, Amr Gomaa, Sarath Sivaprasad, Lea Schönherr, and Mario Fritz. 2023. Llm-deliberation: Evaluating llms with interactive multi-agent negotiation games. arXiv preprint arXiv:2309.17234. [↗](https://arxiv.org/abs/2309.17234)
2. Yoram Bachrach, Richard Everett, Edward Hughes, Angeliki Lazaridou, Joel Z. Leibo, Marc Lanctot, Michael Johanson, Wojciech M. Czarnecki, and Thore Graepel. 2020. Negotiating team formation using deep reinforcement learning. Artificial Intelligence, 288:103356. [find](https://scholar.google.com/scholar?q=Yoram+Bachrach%2C+Richard+Everett%2C+Edward+Hughes%2C+Angeliki+Lazaridou%2C+Joel+Z%2E+Leibo%2C+Marc+Lanctot%2C+Michael+Johanson%2C+Wojciech+M%2E+Czarnecki%2C+and+Thore+Graepel%2E+2020%2E+Negotiating+team+formation+using+deep)
3. Bala M Balachandran and Masoud Mohammadian. 2015. Development of a fuzzy-based multi-agent system for e-commerce settings. Procedia Computer Science, 60:593–602. [find](https://scholar.google.com/scholar?q=Bala+M+Balachandran+and+Masoud+Mohammadian%2E+2015%2E+Development+of+a+fuzzy%2Dbased+multi%2Dagent+system+for+e%2Dcommerce+settings%2E+Procedia+Computer+Science%2C+60%3A593%E2%80%93602%2E)
4. James CL Chow, Leslie Sanders, and Kay Li. 2023. Impact of chatgpt on medical chatbots as a disruptive technology. Frontiers in Artificial Intelligence, 6:1166014. [find](https://scholar.google.com/scholar?q=James+CL+Chow%2C+Leslie+Sanders%2C+and+Kay+Li%2E+2023%2E+Impact+of+chatgpt+on+medical+chatbots+as+a+disruptive+technology%2E+Frontiers+in+Artificial+Intelligence%2C+6%3A1166014%2E)
5. Yao Fu, Hao Peng, Tushar Khot, and Mirella Lapata. 2023. [Improving language model negotiation with self-play and in-context learning from ai feedback](https://md.research.elamur.ai/free/en/papers/fu-2023-improving-language-model-negotiation-self/).
6. He He, Derek Chen, Anusha Balakrishnan, and Percy Liang. 2018. [Decoupling strategy and generation in negotiation dialogues](https://md.research.elamur.ai/free/en/papers/he-2018-decoupling-strategy-generation-negotiation-dialogues/). In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2333–2343.
7. Mohamed Jawahar Hussain. 2014. Parallel processing and intercommunication of multiple negotiation processes in e-commerce for optimal business. In 2014 Fourth International Conference on Communication Systems and Network Technologies, pages 583–586. IEEE. [find](https://scholar.google.com/scholar?q=Mohamed+Jawahar+Hussain%2E+2014%2E+Parallel+processing+and+intercommunication+of+multiple+negotiation+processes+in+e%2Dcommerce+for+optimal+business%2E+In+2014+Fourth+International+Conference+on+Communication)
8. Nicholas R Jennings, Peyman Faratin, Alessio R Lomuscio, Simon Parsons, Carles Sierra, and Michael Wooldridge. 2001. Automated negotiation: prospects, methods and challenges. International Journal of Group Decision and Negotiation, 10(2):199–215. [find](https://scholar.google.com/scholar?q=Nicholas+R+Jennings%2C+Peyman+Faratin%2C+Alessio+R+Lomuscio%2C+Simon+Parsons%2C+Carles+Sierra%2C+and+Michael+Wooldridge%2E+2001%2E+Automated+negotiation%3A+prospects%2C+methods+and+challenges%2E+International+Journal+of+)
9. Gaurav Koley and Shrisha Rao. 2018. Adaptive human-agent multi-issue bilateral negotiation using the thomas-kilmann conflict mode instrument. In 2018 IEEE/ACM 22nd International Symposium on Distributed Simulation and Real Time Applications (DS-RT), pages 1–5. IEEE. [find](https://scholar.google.com/scholar?q=Gaurav+Koley+and+Shrisha+Rao%2E+2018%2E+Adaptive+human%2Dagent+multi%2Dissue+bilateral+negotiation+using+the+thomas%2Dkilmann+conflict+mode+instrument%2E+In+2018+IEEE%2FACM+22nd+International+Symposium+on+Distribut)
10. János Kramár, Tom Eccles, Ian Gemp, Andrea Tacchetti, Kevin McKee, Mateusz Malinowski, Thore Graepel, and Yoram Bachrach. 2022. Negotiation and honesty in artificial intelligence methods for the board game of diplomacy. Nature Communications, 13. [find](https://scholar.google.com/scholar?q=J%C3%A1nos+Kram%C3%A1r%2C+Tom+Eccles%2C+Ian+Gemp%2C+Andrea+Tacchetti%2C+Kevin+McKee%2C+Mateusz+Malinowski%2C+Thore+Graepel%2C+and+Yoram+Bachrach%2E+2022%2E+Negotiation+and+honesty+in+artificial+intelligence+methods+for+the+board)
11. Mike Lewis, Denis Yarats, Yann Dauphin, Devi Parikh, and Dhruv Batra. 2017. [Deal or no deal? end-to-end learning of negotiation dialogues](https://md.research.elamur.ai/free/en/papers/lewis-2017-deal-or-no-deal-end/). In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2443–2453, Copenhagen, Denmark. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/d17-1259)
12. Kshitij Mishra, Azlaan Mustafa Samad, Palak Totala, and Asif Ekbal. 2022. [Pepds: A polite and empathetic persuasive dialogue system for charity donation](https://md.research.elamur.ai/free/en/papers/mishra-2022-pepds-polite-empathetic-persuasive-dialogue/). In Proceedings of the 29th International Conference on Computational Linguistics, pages 424–440.
13. Brian Plüss and Paul Piwek. 2016. Measuring non-cooperation in dialogue. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1925– 1936, Osaka, Japan. [find](https://scholar.google.com/scholar?q=Brian+Pl%C3%BCss+and+Paul+Piwek%2E+2016%2E+Measuring+non%2Dcooperation+in+dialogue%2E+In+Proceedings+of+COLING+2016%2C+the+26th+International+Conference+on+Computational+Linguistics%3A+Technical+Papers%2C+pages+1925%E2%80%93+19)
14. The COLING 2016 Organizing Committee. [find](https://scholar.google.com/scholar?q=The+COLING+2016+Organizing+Committee%2E)
15. Aritra Raut, Abhisek Tiwari, Subrata Das, Sriparna Saha, Anutosh Maitra, Roshni Ramnani, and Shubhashis Sengupta. 2023. Reinforcing personalized persuasion in task-oriented virtual sales assistant. Plos one, 18(1):e0275750. [find](https://scholar.google.com/scholar?q=Aritra+Raut%2C+Abhisek+Tiwari%2C+Subrata+Das%2C+Sriparna+Saha%2C+Anutosh+Maitra%2C+Roshni+Ramnani%2C+and+Shubhashis+Sengupta%2E+2023%2E+Reinforcing+personalized+persuasion+in+task%2Doriented+virtual+sales+assistant%2E+Pl)
16. Andreas Schmid, Oliver Kronberger, Niklas Vonderach, and Mareike Schoop. 2021. Are you for real? a negotiation bot for electronic negotiations. [find](https://scholar.google.com/scholar?q=Andreas+Schmid%2C+Oliver+Kronberger%2C+Niklas+Vonderach%2C+and+Mareike+Schoop%2E+2021%2E+Are+you+for+real%3F+a+negotiation+bot+for+electronic+negotiations%2E)
17. David Traum, Stacy Marsella, Jonathan Gratch, Jina Lee, and Arno Hartholt. 2008. Multi-party, multi-issue, multi-strategy negotiation for multi-modal virtual agents. pages 117–130. [find](https://scholar.google.com/scholar?q=David+Traum%2C+Stacy+Marsella%2C+Jonathan+Gratch%2C+Jina+Lee%2C+and+Arno+Hartholt%2E+2008%2E+Multi%2Dparty%2C+multi%2Dissue%2C+multi%2Dstrategy+negotiation+for+multi%2Dmodal+virtual+agents%2E+pages+117%E2%80%93130%2E)
18. Xuewei Wang, Weiyan Shi, Richard Kim, Yoojung Oh, Sijia Yang, Jingwen Zhang, and Zhou Yu. 2019. [Persuasion for good: Towards a personalized persuasive dialogue system for social good](https://md.research.elamur.ai/free/en/papers/wang-2019-persuasion-good-personalized-persuasive-dialogue/). In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5635–5649, Florence, Italy. Association for Computational Linguistics. [↗](https://doi.org/10.18653/v1/p19-1566)
19. Yiheng Zhou, He He, Alan W Black, and Yulia Tsvetkov. 2019. A dynamic strategy coach for effective negotiation. In Proceedings of the 20th Annual SIGdial Meeting on Discourse and Dialogue, pages 367–378, Stockholm, Sweden. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Yiheng+Zhou%2C+He+He%2C+Alan+W+Black%2C+and+Yulia+Tsvetkov%2E+2019%2E+A+dynamic+strategy+coach+for+effective+negotiation%2E+In+Proceedings+of+the+20th+Annual+SIGdial+Meeting+on+Discourse+and+Dialogue%2C+pages+367%E2%80%933)

*Updated 2 November 2026*