Personalized Emphasis Framing for Persuasive Message Generation

Tao Ding, Shimei PanEMNLP 2016 · DOI

Introduction

Persuasion is most effective when it is tailored to the audience. However, most existing Natural Language Generation (NLG) systems rely on domain-specific user characteristics, such as a user’s smoking habits to tailor a cessation message. This reliance makes systems difficult to scale or transfer to new topics. This paper addresses the challenge of emphasis framing—the act of focusing on specific aspects of an issue (e.g., framing a car by its "safety" vs. its "performance") to appeal to a specific individual. The goal is to build a generalizable model that uses domain-independent psychological traits to decide which information will be most persuasive to a particular person.

Psychological Framework and Trait Acquisition

The researchers developed a multi-stage approach to bridge the gap between psychology and message generation using two established psychometric models:

  1. Big Five Personality Traits: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism.
  2. Schwartz’s Basic Human Values: Including dimensions like Self-transcendence (caring for others/nature) and Conservation (valuing tradition and stability).

By using these universal traits, the system can predict preferences in specific commercial domains, such as automobile purchasing, without requiring prior domain-specific data from the user.

The Cascade Content Selection Model

The core of the proposed system is the Cascade Content Selection Model. When presented with a product and a user profile, the model follows a three-step ranking process:

  1. Objective Ranking: It first ranks aspects based on their objective value (e.g., a car with "excellent" fuel economy ranks higher than one with "average" fuel economy).
  2. Trait-Based Adjustment: It adjusts these ranks using regression models that map psychological traits to aspect preferences.
  3. Interaction Refinement: It further refines the ranks using "interaction rules" derived from user clusters (e.g., how a combination of high Extraversion and high Neuroticism affects preference).

The model finally outputs the top n aspects to be included in the generated persuasive message.

Key Insights and Pattern Discovery

Statistical analysis revealed that domain-independent traits are highly predictive of specific preferences. For example, individuals scoring high in "Conservation" consistently prioritize "Safety." Interestingly, the study found that "Self-transcendence" was a strong predictor for prioritizing "Fuel Economy" but was negatively correlated with "Style," suggesting that highlighting aesthetics might actually backfire for certain value-driven individuals.

Regression coefficients mapping psychological traits to car aspect preferences.

Furthermore, the study found that non-expert humans are surprisingly poor at manual personalized framing. In an "Author Framing Dataset" collected via Amazon Mechanical Turk, human authors failed to produce statistically significant persuasive effects when trying to tailor messages manually. This justifies the need for an algorithmic approach to capture the subtle logic of psychological framing.

Experimental Results

The authors evaluated the Cascade Model against a baseline that only considered the objective values of the car aspects (i.e., always highlighting the "best" features regardless of the user).

Comparison of Top-1 and Top-3 agreement between the Cascade model and the baseline.

As shown in the results, the Cascade Model significantly outperformed the baseline:

  • Top-1 Agreement: The model correctly predicted the user's primary preference 62% of the time, compared to 54% for the baseline.
  • Top-3 Agreement: When selecting the top three aspects for a message, the model achieved 87% agreement with user preferences, a massive improvement over the baseline’s 46%.

Conclusion and Future Work

This research demonstrates that universal psychological profiles can effectively guide content selection in persuasive messaging. While the traits used are domain-independent, the current mappings were learned specifically for the automobile domain. Future work will explore how these mappings adapt to other areas like healthcare or politics. Additionally, the researchers plan to integrate this content selection model with linguistic realization techniques to ensure the tone and style of the generated text also align with the user's personality.

Summary updated

BibTeX
@inproceedings{ding2016personalized,
  author = {Tao Ding and Shimei Pan},
  title = {{Personalized Emphasis Framing for Persuasive Message Generation}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2016},
  doi = {10.18653/v1/d16-1150},
}
Figure