Enhancing Persuasive Dialogue Agents by Synthesizing Cross-Disciplinary Communication Strategies

1. Problem Statement

Current research on persuasive dialogue agents (PDAs) faces a significant limitation: strategic narrowness. Most existing systems rely on a small, predefined set of annotated strategies (e.g., the 10 strategies in the Persuasion for Good or P4G dataset). These approaches often fail to capture the complexity of real-world interactions and overlook a vast repository of proven techniques from social psychology, behavioral economics, and communication theory. Consequently, current agents struggle to effectively persuade individuals with low initial intent or to adapt to diverse, multi-domain scenarios.

2. Methodology

The authors propose a cross-disciplinary framework that synthesizes foundational principles from multiple fields to create a robust, extended set of persuasive strategies.

A. Cross-Disciplinary Strategy Synthesis

The framework expands the standard 10 P4G strategies into a comprehensive set of 31 strategies, categorized into seven distinct groups:

1. Inquiries: Gathering information (e.g., source-related, task-related, personal inquiries).
2. Persuasion Routes: Based on the Elaboration Likelihood Model (ELM), utilizing Central Route (Logical appeal) for durable change and Peripheral Route (Emotional appeal) for quick influence.
3. Trust & Credibility: Building trust via credibility appeals, authority, social proof, and consistency.
4. Action Facilitation: Techniques like "Foot-in-the-door," "Door-in-the-face," and "Reciprocity."
5. Information Presentation: Utilizing Framing, Scarcity, and Time Pressure (from behavioral economics) to alter decision-making.
6. Personalization: Tailoring messages to individual values and stories.
7. Follow-up: Maintaining engagement and reinforcing commitment.

B. Implementation: Proactive Chain-of-Thought (ProCoT)

To operationalize these strategies without expensive retraining, the authors employ a prompting-based approach using Large Language Models (LLMs), specifically GPT-4o.

• Mechanism: The system uses a Proactive Chain-of-Thought (ProCoT) scheme. The LLM is instructed to:
  1. Analyze the dialogue history and context.
  2. Infer the most suitable strategy from the extended set.
  3. Generate a response based on that strategy.
• Variations: The study compares agents using only P4G strategies (ProCoT-p4g) against those using the extended set (ProCoT-rich). It also tests the impact of including strategy descriptions in the prompt (-desc variants).

C. Evaluation Framework

The framework was validated on two distinct datasets:

1. P4G (Persuasion for Good): A single-domain dataset focused on charity donations.
   • Metrics: Success Rate (SR), Average Turns (AT), and Average Intention Improvement (AII) (measuring attitude shifts even in failed dialogues).
   • Simulation: Used an LLM-based persuadee simulator with specific personas (Big Five personality, decision-making style) to evaluate intent shifts on a 5-level scale.
2. DailyPersuasion: A large-scale, multi-domain dataset (35 domains, 13k scenarios).
   • Metrics: Win-rate against baselines using an LLM-as-a-judge approach.

3. Key Contributions

• Extended Strategy Set: The creation of a 31-strategy framework that bridges the gap between AI research and established theories in psychology and economics.
• Granular Intention Modeling: Introduction of the Average Intention Improvement (AII) metric, which quantifies how much an agent shifts a user's attitude even if the conversation does not end in immediate success.
• Strategy-Context Analysis: A detailed analysis of which strategies work best for specific user states (e.g., "Credibility appeal" works best for skeptical users, while "Foot-in-the-door" works for those with high initial intent).
• Ethical Framework: A proposal for a multi-layered safety protocol (Alignment-based filtering, Harmfulness detection, and Human-in-the-Loop) to prevent the misuse of persuasive techniques for manipulation.

4. Experimental Results

The proposed framework, particularly ProCoT-rich-desc (extended strategies with descriptions), demonstrated superior performance:

• Success Rates (P4G):
  • Achieved the highest success rate (83.3%) compared to baselines (Simple: 55.3%, ProCoT-p4g: 76.0%).
  • Low-Intent Persuasion: It was significantly more effective at persuading users with low initial intent (Levels 4 and 5), achieving success rates where baselines failed.
• Intention Improvement (AII):
  • ProCoT-rich-desc achieved the highest AII (0.480), indicating it successfully shifted the attitudes of users even when they did not donate.
  • Notably, it achieved an AII of 1.100 for users with "Initially Reluctant" (Level 4) intent, showing a massive positive shift.
• Generalizability (DailyPersuasion):
  • Outperformed baselines across diverse domains (Win rate of 54.4% vs. Simple and 35.1% vs. ProCoT-p4g).
  • Demonstrated robustness in 35 different domains, though performance dipped in the "Law" domain due to a lack of specific tailoring.
• Human Evaluation:
  • Annotators preferred ProCoT-rich-desc over ProCoT-p4g in 72.5% of cases regarding persuasiveness.
  • Realism ratings were high (3.73/5), though some feedback noted issues with repetitive phrasing or lack of empathy in specific turns.
• Efficiency vs. Cost:
  • While ProCoT-rich-desc had slightly longer response times and higher token usage (and thus cost) due to the Chain-of-Thought reasoning, the trade-off resulted in significantly higher persuasion efficacy.

5. Significance and Future Directions

• Practical Impact: The study proves that integrating cross-disciplinary knowledge into LLM prompting can create agents that are not only more persuasive but also capable of handling resistant users—a critical challenge in real-world applications like healthcare, sales, and social good.
• Granularity Matters: The results highlight that treating "low intent" as a monolithic group is insufficient; fine-grained segmentation (e.g., distinguishing between "Negative Reaction" and "Explicit Non-donor") allows for more targeted and effective strategies.
• Ethical Responsibility: The authors emphasize that the ability to shift low-intent users is a double-edged sword. They argue for context-aware deployment and robust safety filters to prevent manipulation, suggesting that future systems must balance efficacy with ethical safeguards.
• Future Work: The authors plan to validate these findings with real human participants, explore the combination of strategies (e.g., "Door-in-the-face" followed by "Foot-in-the-door"), and refine the framework for specific high-stakes domains like legal and medical advice.

In conclusion, this paper establishes that knowledge-driven, cross-disciplinary strategy synthesis significantly outperforms data-driven, single-domain approaches in creating effective and adaptable persuasive dialogue agents.