Understanding User Resistance Strategies in Persuasive Conversations

Youzhi Tian, Weiyan Shi, Chen Li, Zhou YuEMNLP 2020 · OpenReview

Abstract

Persuasive dialog systems have various usages, such as donation persuasion and physical exercise persuasion. Previous persuasive dialog systems research mostly focused on analyzing the persuader’s strategies and paid little attention to the persuadee (user). However, understanding and addressing users’ resistance strategies is an essential job of a persuasive dialog system. So, we adopt a preliminary framework on persuasion resistance in psychology and design a fine-grained resistance strategy annotation scheme. We annotate the PersuasionForGood dataset with the scheme. With the enriched annotations, we build a classifier to predict the resistance strategies. Furthermore, we analyze the relationships between persuasion strategies and persuasion resistance strategies. Our work lays the ground for developing a persuasive dialogue system that can understand and address user resistance strategy appropriately.

Introduction

In the field of Natural Language Processing (NLP), building persuasive agents—bots designed to convince users to donate to charity or adhere to health routines—is a rapidly growing area of interest. However, current systems suffer from a significant limitation: they lack a nuanced understanding of user resistance.

Most existing models treat any user pushback as a generic "negative reaction." Whether a user says, "I can't afford it," "I don't trust this charity," or "Let's talk about something else," the system often fails to distinguish between these fundamentally different psychological stances. This lack of granularity makes it nearly impossible for a bot to respond with an appropriate counter-strategy. To build truly intelligent persuasive systems, we must move beyond viewing resistance as a monolith and begin identifying the specific psychological tactics users employ to remain unpersuaded.

A Taxonomy of Resistance

To address this gap, the researchers grounded their work in a psychological framework by Fransen et al. (2015), which categorizes resistance into four primary mechanisms: Avoidance, Contesting, Biased Processing, and Empowerment.

The hierarchical taxonomy of resistance strategies used to categorize user responses.

From this foundation, the authors developed a fine-grained annotation scheme identifying six strategies that frequently appear in real-world persuasive dialogues:

  1. Avoidance: Shifting the topic or ignoring the persuasive attempt.
  2. Self-assertion: Expressing confidence in one's current stance (e.g., "I'm sure I won't change my mind").
  3. Weighting attributes: Focusing on specific negative attributes or concerns (e.g., questioning fund allocation).
  4. Contesting content: Challenging the accuracy or relevance of the persuader's claims.
  5. Attitude bolstering: Reaffirming existing beliefs to resist change.
  6. Contesting strategy: Challenging the persuader’s motives or the way they are communicating.

Methodology and Dataset

The study utilized the PERSUASIONFORGOOD dataset, which consists of 1,017 dialogues where a persuader attempts to convince a persuadee to donate to the charity Save the Children.

Experts annotated 447 specific resistant utterances within this dataset. The researchers then trained a Recurrent Convolutional Neural Network (RCNN) to automatically detect these strategies. Due to data imbalance—where some strategies like "Contesting strategy" were extremely rare—the categories were grouped into four main buckets for the machine learning model to ensure more robust training.

Key Insights: Resistance as a Conversion Killer

The analysis revealed that the presence of resistance is a primary indicator of dialogue failure. In conversations where no resistance was detected, the donation rate was a high 54%. However, once a user employed even a single resistance strategy, the donation rate plummeted to 22%.

Donation rates across different resistance categories, highlighting the impact of specific strategies on persuasion success.

Crucially, not all forms of resistance are equally insurmountable. The study found that Self-assertion is the most difficult to overcome, resulting in the lowest donation rate (15%). In contrast, Weighting attributes—where a user focuses on a specific concern—resulted in a higher donation rate of 26%. This suggests that while a flat refusal (Self-assertion) may be a dead end, a user who questions specific attributes is still engaged and potentially open to persuasion if the bot can provide logical evidence or a credibility appeal to address that specific concern.

Results and Future Directions

The RCNN classifier achieved an accuracy of 67.34% and a macro F1 score of 0.52. While there is room for improvement, the results demonstrate that psychological resistance strategies possess distinct linguistic patterns that machines can learn to identify. A common point of confusion for the model was distinguishing between "Weighting attributes" and "Avoidance," likely because users often use avoidant language when they find a specific attribute of the persuasion inconsistent with their beliefs.

Future work will focus on integrating this resistance detector into live systems to test if "strategy-aware" responses can effectively pivot a conversation and increase success rates. Additionally, these insights can be used to build more realistic user simulators, providing a safer and more effective environment for training persuasive AI.

Summary updated

Abstract

Persuasive dialog systems have various usages, such as donation persuasion and physical exercise persuasion. Previous persuasive dialog systems research mostly focused on analyzing the persuader’s strategies and paid little attention to the persuadee (user). However, understanding and addressing users’ resistance strategies is an essential job of a persuasive dialog system. So, we adopt a preliminary framework on persuasion resistance in psychology and design a fine-grained resistance strategy annotation scheme. We annotate the PersuasionForGood dataset with the scheme. With the enriched annotations, we build a classifier to predict the resistance strategies. Furthermore, we analyze the relationships between persuasion strategies and persuasion resistance strategies. Our work lays the ground for developing a persuasive dialogue system that can understand and address user resistance strategy appropriately. The code and data will be released.
BibTeX
@inproceedings{tian2020understanding,
  author = {Youzhi Tian and Weiyan Shi and Chen Li and Zhou Yu},
  title = {{Understanding User Resistance Strategies in Persuasive Conversations}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=2_jFIRLUoiH},
}
Figure