Enhancing Persuasive Dialogue Agents by Synthesizing Cross-Disciplinary Communication Strategies
Enhanced persuasive agents using cross-disciplinary strategies, achieving higher success rates, especially for low-intent users.
Key Findings
Methodology
The study introduces a cross-disciplinary framework synthesizing strategies from social psychology, behavioral economics, and communication theory. It employs ProCoT prompting with LLMs to dynamically generate persuasive dialogues using an extended set of 31 strategies.
Key Results
- On the P4G dataset, the ProCoT-rich-desc model achieved an 83.3% persuasion success rate, a 7.3% improvement over the ProCoT-p4g baseline. It excelled in persuading low-intent users (initial intent levels 4 and 5).
- On the DailyPersuasion dataset, ProCoT-rich-desc achieved a win rate of 54.4% against Simple and 35.1% against ProCoT-p4g, demonstrating robust cross-domain generalizability.
- In failed dialogues, ProCoT-rich-desc achieved an Average Intention Improvement (AII) of 0.48, outperforming baselines and showing potential for shifting low-intent users' attitudes.
Significance
This work addresses the limitations of existing persuasive agents by introducing a richer, cross-disciplinary strategy set. The results show significant improvements in persuasion success rates and the ability to influence low-intent users, with applications in healthcare, sales, and social good.
Technical Contribution
Key contributions include: 1) a comprehensive set of 31 cross-disciplinary strategies; 2) the ProCoT prompting technique for dynamic strategy selection and response generation; 3) stricter evaluation metrics, including intention improvement and cross-domain generalizability.
Novelty
This is the first framework to systematically integrate strategies from multiple disciplines into persuasive dialogue agents. It significantly extends the P4G strategy set and introduces ProCoT for efficient, dynamic generation.
Limitations
- Limited success with users having the lowest intent (level 5), highlighting the need for further strategy refinement.
- Generated dialogues sometimes lack diversity, leading to repetitive expressions and reduced naturalness.
- High computational costs may hinder scalability in real-world applications.
Future Work
Future research could explore more efficient strategy generation methods, reduce computational costs, and improve strategy diversity to enhance user experience and scalability.
AI Executive Summary
Existing persuasive dialogue agents often rely on predefined strategies, limiting their ability to handle complex real-world interactions. This study introduces a cross-disciplinary framework that integrates strategies from social psychology, behavioral economics, and communication theory. Using ProCoT prompting, the framework dynamically generates persuasive dialogues with an extended set of 31 strategies, far surpassing the 10 strategies in the P4G dataset.
Experiments on the P4G and DailyPersuasion datasets demonstrate the framework's effectiveness. On P4G, the ProCoT-rich-desc model achieved an 83.3% success rate, excelling with low-intent users. On DailyPersuasion, it outperformed baselines across diverse domains, showcasing strong generalizability. Even in failed dialogues, it significantly improved users' donation intentions.
However, challenges remain, including limited success with the lowest-intent users, repetitive dialogue patterns, and high computational costs. Future work will focus on optimizing strategies, enhancing dialogue diversity, and expanding real-world applications in healthcare, sales, and social initiatives.
Deep Analysis
Background
Advances in LLMs have improved dialogue systems' reasoning and interaction capabilities, driving research into persuasive agents. However, most existing systems rely on limited predefined strategies, such as the 10 strategies in the P4G dataset, failing to leverage rich insights from psychology and economics. This limits their effectiveness, especially with low-intent users.
Core Problem
The core challenge lies in the narrow scope and lack of generalizability of current persuasive strategies. Many systems are domain-specific and fail to address the nuanced dynamics of real-world interactions, particularly with users who are initially resistant to persuasion.
Innovation
Key innovations include: 1) a cross-disciplinary framework integrating ELM and HSM models with behavioral economics strategies; 2) the ProCoT prompting method for dynamic strategy selection; 3) an expanded set of 31 strategies, addressing gaps in existing datasets like P4G.
Methodology
- �� Developed a cross-disciplinary framework integrating ELM (central/peripheral routes) and behavioral economics strategies like scarcity and framing.
- �� Used ProCoT prompting to dynamically generate persuasive responses by analyzing dialogue history and selecting optimal strategies.
- �� Validated the framework on P4G and DailyPersuasion datasets, measuring success rates, intention improvement, and efficiency.
Experiments
Experiments used P4G and DailyPersuasion datasets. P4G (300 samples) evaluated success rate (SR), average turns (AT), and intention improvement (AII). DailyPersuasion (1,000 samples) assessed cross-domain generalizability. Baselines included Simple and ProCoT-p4g.
Results
On P4G, ProCoT-rich-desc achieved an 83.3% success rate, outperforming baselines. On DailyPersuasion, it achieved a win rate of 54.4% (vs. Simple) and 35.1% (vs. ProCoT-p4g), demonstrating cross-domain robustness.
Applications
Applications include healthcare interventions (e.g., promoting healthy habits), sales support (e.g., increasing purchase intent), and social initiatives like charity fundraising.
Limitations & Outlook
Limitations include: 1) limited success with lowest-intent users; 2) repetitive dialogue patterns; 3) high computational costs, limiting scalability.
Plain Language Accessible to non-experts
Imagine you're trying to convince a friend to donate to a charity. You could use logic (explaining how the money will be used) or emotions (sharing a touching story). This research acts like a 'persuasion expert,' teaching AI to pick the best strategy dynamically. For example, if your friend is hesitant, the AI might suggest donating a small amount first ('foot in the door') or emphasize urgency ('time pressure'). Experiments show this approach works especially well for people who are initially resistant.
ELI14 Explained like you're 14
Think of it like a game where you need to convince NPCs to donate to save animals. You can use different strategies, like telling a heartwarming story or showing how their donation helps. This research is like a cheat code—it teaches AI 31 strategies (way more than before!) to win. It’s especially good at convincing NPCs who don’t want to donate at first. Cool, right? But sometimes the AI talks too much or repeats itself, so there’s room to improve.
Glossary
ELM (Elaboration Likelihood Model)
A psychology model distinguishing central (logic-based) and peripheral (emotion-based) persuasion routes.
Used to design diverse persuasive strategies.
Behavioral Economics
A field studying how psychological factors influence decision-making.
Informs strategies like framing and scarcity.
ProCoT Prompting
A method using Chain-of-Thought reasoning to dynamically generate dialogue responses.
Implemented for strategy selection in this study.
P4G Dataset
A dataset for persuasion research with 10 predefined strategies.
Used as a baseline for evaluation.
AII (Average Intention Improvement)
Measures changes in user intent during failed dialogues.
Evaluates effectiveness with low-intent users.
Open Questions Unanswered questions from this research
- 1 How can we further improve success rates for lowest-intent users (level 5)?
- 2 What methods can reduce computational costs for real-world deployment?
- 3 How can strategy diversity and dialogue naturalness be enhanced?
Applications
Immediate Applications
Healthcare Interventions
Encourage healthy behaviors like quitting smoking or exercising through dynamic persuasion.
Sales Support
Boost customer purchase intent by optimizing persuasive dialogues.
Long-term Vision
Social Impact Optimization
Use AI persuasion to drive social good, such as increasing charity donations or promoting sustainability.
Abstract
Current approaches to developing persuasive dialogue agents often rely on a limited set of predefined persuasive strategies that fail to capture the complexity of real-world interactions. We applied a cross-disciplinary approach to develop a framework for designing persuasive dialogue agents that draws on proven strategies from social psychology, behavioral economics, and communication theory. We validated our proposed framework through experiments on two distinct datasets: the Persuasion for Good dataset, which represents a specific in-domain scenario, and the DailyPersuasion dataset, which encompasses a wide range of scenarios. The proposed framework achieved strong results for both datasets and demonstrated notable improvement in the persuasion success rate as well as promising generalizability. Notably, the proposed framework also excelled at persuading individuals with initially low intent, which addresses a critical challenge for persuasive dialogue agents.