Can a model recommend the right persuasive move, and explain why?
Recommending persuasive strategies works better, and is easier to trust, when a model knows both speaker and audience and cites evidence from the conversation itself.
Hypotheses.
What the paper argues, written so it can be tested.
- H1
Recommendations that know both the speaker and the audience pick effective persuasive primitives better than recommendations from context alone.
- H2
Recommendations that cite evidence from the conversation are trusted more by the people who use them, without losing accuracy.
The model.
PRISM: the dialogue, its context, the speaker, and the audience go in; recommended primitives with cited reasons come out; then the delivered turn is scored against that advice.
A player lifting a piece mid-game.
Photo: Jent JiangPartial abstract.
PRISM recommends the persuasive primitives, such as emotional appeal or logical coherence, most likely to work in the next turn of a dialogue, then scores the turn once it is said. Its labels come from an oracle model that explains each primitive with evidence from the conversation and the speaker’s and audience’s personas.
PRISM: Personalized Recommendation of In-Context Strategies with Grounded Explanations for Effective Dialogue
Key sources
- Ning, L., et al. (2024). USER-LLM: Efficient LLM contextualization with user embeddings.
- Su, K., et al. (2024). REGEN: A dataset and benchmarks with natural language critiques and narratives. NAACL.
How we’ll test it.
No results yet. This is the plan: the data, the measure, and what would count as support.
The data
Dialogues with known outcomes, such as debates, labeled by an oracle model with counterfactuals, and held-out dialogues it never saw.
The measure
Accuracy on held-out dialogues against context-only and non-personalized baselines, and trust ratings from people shown explained and unexplained advice.
What would count as support
The personalized model beats both baselines on held-out dialogues, and explanations raise trust without lowering accuracy.
How it connects to Pulp.
TrueAura scores persuasive resonance against an objective.
Also in Persuasion
- Why do parasocial fandoms fall apart?
J.Q., Boston University
- How does crime news shape what people fear, and what they support?
A.V., CUNY Graduate Center
Doing work like this?
Students from any school or field, and mentors from academia or industry: get in touch.