TLDR: A new research paper introduces the Personality-driven Argumentation-based Negotiation Dialogue Generation (PAN-DG) task and the PACT dataset for tourism. This work focuses on creating AI negotiation systems that understand and adapt to a traveler’s unique personality, preferences, and buying style. By using LLMs to generate and filter a large dataset of negotiation dialogues, and then fine-tuning models with this data, the researchers demonstrate significant improvements in generating personalized, rational, and factually accurate responses for complex travel planning negotiations.
The tourism industry is experiencing rapid growth, and with it comes a demand for more sophisticated dialogue systems to help travelers plan and book their trips. Unlike simple transactions, planning a vacation involves complex negotiations, considering diverse user preferences, budget constraints, and various interdependent factors like price, accommodation, and destination. Current Large Language Models (LLMs) often struggle with these intricate scenarios.
A new research paper, “We Argue to Agree: Towards Personality-Driven Argumentation-Based Negotiation Dialogue Systems for Tourism”, introduces a novel approach to address this challenge. The paper proposes the Personality-driven Argumentation-based Negotiation Dialogue Generation (PAN-DG) task, aiming to create negotiation dialogue systems that are not only effective in resolving conflicts but also highly personalized to individual travelers’ preferences and negotiation styles.
Understanding the PAN-DG Task
The PAN-DG task is designed to generate negotiation dialogues that are tailored to an individual’s personality while incorporating robust argumentation techniques. This leads to more personalized, rational, and collaborative negotiation outcomes. The task is broken down into three core sub-tasks:
- Personality Recognition: Identifying the personality profiles of the participants based on the dialogue context.
- Dialog Act Prediction: Inferring the next likely action or intention in the conversation.
- Response Generation: Creating appropriate and relevant responses from the travel agent.
Introducing PACT: A New Dataset for Tourism Negotiations
To support the PAN-DG task, the researchers developed PACT (Personality-driven Argumentation-based negotiation Conversations for Tourism). Recognizing the difficulty and time involved in manually creating high-quality conversational datasets, PACT was automatically generated using LLMs, with human oversight through a meticulous three-stage pipeline:
- PACT Generation: The Gemini-1.5 LLM was used to generate negotiation dialogues, specifically tailored to diverse personality profiles.
- PACT Filtering: A GPT-4-based mixture-of-experts approach was employed to ensure the high quality of the generated dialogues, filtering out erroneous or inconsistent conversations.
- PACT Quality Assessment: Both automatic and manual evaluations were conducted to rigorously assess the dialogue quality, ensuring lexical and semantic diversity, coherence, and naturalness.
A key innovation in PACT is the definition of three distinct personality profiles, which go beyond generic traits to capture domain-specific negotiation behaviors:
- Argumentation Profile: Describes how individuals negotiate, persuade, accept, or reject offers (e.g., Agreeable, Disagreeable, Open-minded, Argumentative).
- Preference Profile: Captures specific trip preferences relevant to tourism (e.g., Culture Creature, Thrill Seeker, Beach Lover).
- Buying Style Profile: Characterizes purchasing behaviors and motivations (e.g., Quality-concerned, Budget-concerned, Budget-&-Quality-concerned).
The dataset also incorporates 23 ABN-specific dialogue acts, categorized into negotiation, argumentation, and general acts, to accurately reflect the interactive dynamics of real-world negotiations.
Also Read:
- AI and Language Models Streamline Complex Risk Negotiations for Global Challenges
- Simulating Human Perspective Taking in AI: A Developmental Approach
Experimental Results and Impact
The researchers conducted extensive experiments using various LLMs, including LLaMA models, Mistral, Vicuna, and GPT-4.1-mini. They compared several setups, from pre-trained models to those fine-tuned on PACT, and a specialized setup (FT-Ours) that integrated background package knowledge, personality, and dialog act information.
The results were compelling: models fine-tuned on PACT showed significant improvements over pre-trained models and even outperformed models fine-tuned on general human-human negotiation datasets. The FT-Ours approach, in particular, demonstrated the highest gains across all sub-tasks, leading to substantial improvements in response quality, dialogue coherence, and factual accuracy. This setup significantly reduced perplexity, reflecting enhanced model confidence and fluency, and drastically improved personality and dialog act accuracy.
These findings highlight that incorporating domain-specific knowledge, personality profiles, and structured dialog acts is crucial for developing advanced, personality-driven argumentation-based negotiation dialogue systems. The PACT dataset and the proposed methodology lay a strong foundation for future research, enabling the creation of more robust and scalable AI agents capable of handling the complexities of personalized tourism negotiations.
