Persuasion for Good: Towards a Personalized Persuasive Dialogue System for Social Good
This study introduces a deep learning-based personalized persuasive dialogue system, achieving 59.6% F1 on a 300-sample annotated dataset, advancing social good applications.
Key Findings
Methodology
The research collected a large-scale human-human persuasion dialogue dataset comprising 1017 conversations, employing content analysis to develop a taxonomy of ten persuasion strategies. A hybrid RCNN model was designed, integrating sentence embeddings, context features, and sentence-level cues. Inputs included pre-trained FastText word vectors, bidirectional LSTM encodings for sentences and context, sentiment features via VADER, character-level embeddings from a multiplicative LSTM trained on Amazon reviews, and positional embeddings. The model was trained with five-fold cross-validation, achieving 74.8% accuracy and 59.6% macro F1. Additionally, the study analyzed the relationship between participants’ demographic and psychological backgrounds—such as Big Five traits, moral foundations, Schwartz values, and decision styles—and their donation behaviors, providing insights for personalized strategy recommendations.
Key Results
- The hybrid RCNN model outperformed baseline models (BLSTM and CNN), reaching 74.8% accuracy and 59.6% F1, demonstrating the effectiveness of multi-feature fusion. Ablation studies confirmed that sentence embeddings and sentiment features contributed most significantly to performance gains.
- Analysis of strategy distribution revealed temporal and contextual patterns, such as credibility appeals being more frequent at dialogue start, while donation information appeared predominantly in later turns. Personal background influenced strategy effectiveness, highlighting the importance of personalization.
- In experiments, incorporating context via RNN-based encoding yielded the best results among different context extraction methods. The model's ability to distinguish strategies was validated through confusion matrix analysis, identifying common misclassifications between Personal story and Emotion appeal, and between Logical and Emotional appeals.
Significance
This work bridges a critical gap in automated social good interventions by integrating social psychology theories with advanced deep learning techniques. It provides a scalable framework for understanding and predicting persuasive behaviors in dialogue, enabling the development of intelligent, personalized systems that can effectively promote charitable donations. Such systems have the potential to transform public engagement strategies, making them more targeted and impactful, ultimately fostering greater societal benefits. The dataset and model serve as valuable resources for further research in AI-driven social influence and ethical considerations in automated persuasion.
Technical Contribution
The core technical innovation lies in the hybrid RCNN architecture that fuses multiple modalities—sentence embeddings, context encodings, sentiment, character features, and positional cues—resulting in superior classification performance. The approach advances the state-of-the-art in multi-strategy detection within dialogue systems, demonstrating how deep neural networks can effectively model complex persuasive behaviors. The comprehensive annotation scheme and large dataset also contribute to establishing standardized benchmarks for future research in persuasive dialogue understanding.
Novelty
This study is the first to systematically classify and analyze ten distinct persuasion strategies in a large-scale human dialogue corpus, incorporating psychological background data for personalized analysis. Unlike prior work limited to single-turn or static texts, it employs a multi-feature deep learning framework tailored for multi-turn conversations, enabling nuanced strategy recognition and personalization. The integration of social science insights with cutting-edge NLP models marks a significant step forward in AI-powered social influence research.
Limitations
- The model's performance diminishes in long, semantically complex dialogues, especially when strategies like Personal story and Emotion appeal are confounded, indicating the need for more sophisticated context encoding methods.
- Data collection via AMT may introduce sampling bias, limiting generalization to broader populations. Future work should include diverse data sources to enhance robustness.
- Strategy classification at the sentence level may overlook multi-strategy overlaps within turns; multi-label approaches could improve granularity and accuracy.
Future Work
Future directions include integrating multimodal data such as speech and facial expressions to enrich context understanding, developing personalized strategy recommendation systems based on psychological profiles, and deploying the framework in real-world social platforms for field testing. Additionally, exploring reinforcement learning to adapt strategies dynamically during conversations could further enhance system effectiveness, ultimately leading to scalable, ethical AI agents that promote social good.
AI Executive Summary
In contemporary society, leveraging artificial intelligence to enhance social good initiatives has become a compelling research frontier. Traditional methods of charitable outreach rely heavily on static messaging and manual engagement, which often lack personalization and fail to sustain public interest. With the advent of advanced dialogue systems and deep learning, there is now an unprecedented opportunity to create intelligent agents capable of engaging individuals in persuasive conversations tailored to their psychological and demographic backgrounds.
This study addresses the challenge of designing a socially responsible, personalized persuasive dialogue system by constructing a large-scale dataset of human-human persuasion interactions. The dataset, named PERSUASIONFORGOOD, comprises 1017 dialogues collected via Amazon Mechanical Turk, where participants were tasked with persuading each other to donate to Save the Children. To systematically analyze these interactions, the researchers developed a comprehensive annotation scheme covering ten persuasion strategies, including logical, emotional, credibility, and inquiry-based tactics. A subset of 300 dialogues was annotated, providing a rich resource for training and evaluating machine learning models.
Building on this foundation, the team designed a hybrid deep neural network model—combining recurrent convolutional neural networks (RCNN)—to classify the strategies employed in each utterance. The model integrates multiple features: sentence embeddings derived from FastText, context encodings via bidirectional LSTMs, sentiment scores from VADER, character-level features from a pre-trained mLSTM, and positional embeddings indicating turn order. This multi-modal approach allowed the model to capture both semantic and pragmatic cues, leading to a classification accuracy of 74.8% and a macro F1 score of 59.6%, outperforming baseline models.
Analysis of the results revealed distinct temporal patterns in strategy deployment, with credibility appeals dominating early dialogue and donation information appearing later. Furthermore, by examining participants’ demographic and psychological data—such as Big Five traits, moral foundations, and value systems—the study uncovered significant variations in strategy effectiveness, emphasizing the importance of personalization. These insights provide a pathway toward developing adaptive dialogue agents that can tailor persuasion tactics based on individual profiles.
The significance of this work extends beyond technical innovation. It demonstrates how integrating social science theories with deep learning can produce systems capable of ethically influencing behavior for social good. The dataset and models serve as valuable benchmarks for future research, fostering advancements in AI-driven social influence, ethical persuasion, and human-AI collaboration. Despite current limitations—such as challenges in modeling complex, multi-strategy utterances and potential biases in data collection—the study paves the way for more sophisticated, context-aware, and ethically aligned persuasive agents.
Looking ahead, future research will explore multimodal data integration, real-world deployment, and reinforcement learning-based adaptation to enhance system robustness and personalization. Ultimately, this work aims to empower AI to contribute positively to society by facilitating meaningful, personalized, and ethically responsible social influence campaigns.
Deep Dive
Abstract
Developing intelligent persuasive conversational agents to change people's opinions and actions for social good is the frontier in advancing the ethical development of automated dialogue systems. To do so, the first step is to understand the intricate organization of strategic disclosures and appeals employed in human persuasion conversations. We designed an online persuasion task where one participant was asked to persuade the other to donate to a specific charity. We collected a large dataset with 1,017 dialogues and annotated emerging persuasion strategies from a subset. Based on the annotation, we built a baseline classifier with context information and sentence-level features to predict the 10 persuasion strategies used in the corpus. Furthermore, to develop an understanding of personalized persuasion processes, we analyzed the relationships between individuals' demographic and psychological backgrounds including personality, morality, value systems, and their willingness for donation. Then, we analyzed which types of persuasion strategies led to a greater amount of donation depending on the individuals' personal backgrounds. This work lays the ground for developing a personalized persuasive dialogue system.
References (20)
Reliability in Content Analysis: Some Common Misconceptions and Recommendations
K. Krippendorff
Learning by Communicating in Natural Language With Conversational Agents
A. Graesser, Haiying Li, Carol M. Forsyth
Convolutional Neural Networks for Sentence Classification
Yoon Kim
VADER: A Parsimonious Rule-Based Model for Sentiment Analysis of Social Media Text
C. J. Hutto, Eric Gilbert
Moral Foundations Theory: The Pragmatic Validity of Moral Pluralism
J. Graham, J. Haidt, S. Koleva et al.
A comparison of the invariance properties of the PVQ-40 and the PVQ-21 to measure human values across German and Polish Samples
Jan Cieciuch, E. Davidov
Mapping the Moral Domain
J. Graham, Brian A. Nosek, J. Haidt et al.
The Psychology of Tailoring-Ingredients in Computer-Tailored Persuasion
A. Dijkstra
A Human-Computer Dialogue System for Educational Debate: A Computational Dialectics Approach
Tangming Yuan, D. Moore, Alec Grierson
Guilt appeals: persuasion knowledge and charitable giving
S. Hibbert, Andrew Smith, Andrea Davies et al.
Advancing Tailored Health Communication: A Persuasion and Message Effects Perspective
B. Rimer, M. Kreuter
Recurrent Convolutional Neural Networks for Text Classification
Siwei Lai, Liheng Xu, Kang Liu et al.
Persuasive technology: using computers to change what we think and do
James Kalbach
The automated design of believable dialogues for animated presentation teams
Elisabeth André, T. Rist, Susanne van Mulken et al.
One size does not fit all: The case for tailoring print materials
M. Kreuter, V. Strecher, B. Glassman
The Pleadings Game
T. Gordon
Modifying Socially-Conscious Behavior: The Foot-in-the-Door Technique
C. Scott
Moral Framing and Charitable Donation : Integrating Exploratory Social Media Analyses and Confirmatory Experimentation
B. Donnellan
Dropout: a simple way to prevent neural networks from overfitting
Nitish Srivastava, Geoffrey E. Hinton, A. Krizhevsky et al.
THE DEVELOPMENT OF MARKERS FOR THE BIG-FIVE FACTOR STRUCTURE
L. R. Goldberg
Cited By (20)
PersuHSG: Adaptive Persuasion Strategy Planning for Dialogue Agents Based on Hierarchical Strategy Graph
How Much Does Persuasion Strategy Matter? LLM-Annotated Evidence from Charitable Donation Dialogues
METRO: Towards Strategy Induction from Expert Dialogue Transcripts for Non-collaborative Dialogues
Enhancing Persuasive Dialogue Agents by Synthesizing Cross-Disciplinary Communication Strategies
When Evidence is Sparse: Weakly Supervised Early Failure Alerting in Dialogs and LLM-Agent Trajectories
Detecting Winning Arguments with Large Language Models and Persuasion Strategies
Personality-Aware Reinforcement Learning for Persuasive Dialogue with LLM-Driven Simulation
One Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RL
A General Highly Accurate Online Planning Method Integrating Large Language Models into Nested Rollout Policy Adaptation for Dialogue Tasks
Cognitive World Model for Progressive BDI/E Trajectory Evaluation of Conversational Agents
Information Dynamics of Language Communication
Planner Policies are Bias Learners: A Comprehensive Analysis of Preference Bias in LLM-based Dialogue Planner Policies
Not My Truce: Personality Differences in AI-Mediated Workplace Negotiation
Effective Engagement by Agents for Sense of Ownership and Self-Determination
Characterizing AI Manipulation Risks in Brazilian YouTube Climate Discourse
UP-NRPA: User Portrait based Nested Rollout Policy Adaptation for Planning with Large Language Models in Goal-oriented Dialogue Systems
Train Yourself as an LLM: Exploring Effects of AI Literacy on Persuasion via Role-playing LLM Training
LLM-in-the-Loop Descriptor Prompt Refinement for Dialogue-Based Big Five Personality Recognition
The Hidden Puppet Master: Predicting Human Belief Change in Manipulative LLM Dialogues
PersuaRL: Reinforcement Learning-Driven Multi-Expert Selection for Persuasive Dialogue Generation in Insurance