Modeling Audience Prior Beliefs in Computational Persuasion

17 concepts11 papers2 proposals

Imagine two people reading the exact same political op-ed. One finishes the article feeling deeply validated, while the other dismisses it entirely as biased nonsense. If argument quality were an objective property inherent only to the text, this divergence would be impossible. Yet, traditional computational argumentation has long suffered from a "content-centric" bias, focusing heavily on syntax, logic, and lexical features while treating the audience as a blank slate.

To build systems that truly understand or predict persuasion, the field has shifted toward audience-centric modeling. The core intuition is simple: persuasion does not happen in a vacuum; it happens in the mind of a listener who already possesses a complex web of experiences, values, and convictions. An argument succeeds not just because its logic is flawless, but because it successfully navigates and aligns with the listener's pre-existing internal state.

The Theoretical Framework: Subjective Belief Bases

How do we mathematically formalize the idea that an argument's strength depends on the listener? The foundation lies in separating objective information from subjective belief states. Early work in reasoning under uncertainty by Bacchus et al. demonstrated that subjective beliefs can be formally integrated into objective inference systems by finding probability distributions that minimize information loss while satisfying new probabilistic constraints.

When applied directly to argumentation, this means argument preference is not an arbitrary structural assignment. As Luo et al. showed, we can calculate how much an agent will trust an argument by measuring the conditional satisfiability of its premises relative to that agent's internal belief base. If an argument relies on premises that conflict with what the listener already holds to be true, its perceived strength plummets.

[Argument Premises] ---> Evaluated against ---> [Audience Belief Base] = Subjective Strength

This perspective is formalized in the Epistemic Argumentation Framework (EAF) introduced by Sakama et al.. While traditional frameworks judge arguments purely by how they structurally attack or defend one another, EAFs incorporate an agent's internal degrees of belief. This explains why a structurally sound argument might still fail to persuade: if the epistemic constraint of the listener is not met, the argument cannot take root.

Empirical Proof: The Dominance of the Listener

Data-driven natural language processing has validated these theoretical frameworks, repeatedly proving that a listener's identity frequently overrides the linguistic quality of a message.

In a study of online debate platforms, Durmus et al. discovered that a reader's prior political and religious ideologies are significantly better predictors of who they perceive as the "winner" of a debate than the actual linguistic quality or style of the arguments presented. Even when tracking dynamic interactions over time, Gao et al. found that while evolving topics and discourse styles shape the trajectory of a debate, deep-seated convictions remain a primary source of resistance to opinion shifts.

Furthermore, the specific state of the listener dictates what kind of language will be effective:

  • The Open vs. Closed Mind: Anand et al. demonstrated that while factual "monologues" are generally robust, specific personality types—particularly those high in Openness and Conscientiousness—are uniquely susceptible to emotional, social-media-style arguments.
  • The Undecided vs. The Opposed: Longpre et al. revealed that undecided audience members are primarily swayed by content-rich, semantic language. Conversely, individuals who actually flip their stance from an opposing side are far more influenced by stylistic cues and rhetorical framing than by raw evidence.

Engineering the Audience: Latent Variable Frameworks

If persuasion depends entirely on the audience, how can computational models predict it without forcing every user to take a lengthy personality and ideology survey? Modern NLP solves this by treating the audience's state as a latent (hidden) variable inferred from behavior or text.

One approach relies on behavioral profiling. Al Khatib et al. demonstrated that a user's interests, prior beliefs, and personality traits can be automatically extracted from their historical posting footprints (such as Reddit histories). By modeling these latent characteristics, their system can predict both how persuasive an argument will be to a specific user and how "stubborn" that user is likely to be.

Another elegant solution borrows from psychometrics. Kornilova et al. adapted Item Response Theory (IRT)—traditionally used to separate a student's hidden ability from a test question's difficulty. In the context of persuasion, their framework mathematically decouples the latent "persuadability" of an audience from the latent "persuasiveness" of a message. This allows the model to capture audience-specific variance and explain why the same message alienates one group while convincing another, all without explicit demographic data.

Item Response Theory (IRT) for Persuasion:
P(Stance Change) = f( Audience Persuadability  -  Message Difficulty/Resistance )

For highly personalized applications, Zymla et al. combined linguistic minimal pairs with Gaussian Process Preference Learning. Their system maps out exactly how individual preferences are shaped by subtle rhetorical choices, such as the use of specific sentiment or attitude verbs (e.g., knowing vs. believing).

As a comprehensive survey of the argument quality landscape by Ivanova et al. concludes, the field is moving decisively toward subjectivity-aware modeling. By treating reader background and prior beliefs as essential latent variables rather than noise, modern computational argumentation models are finally reflecting the human reality of persuasion: strength is truly in the eye of the beholder.

Key Foundational Work

Generating New Beliefs From Old (1994) — contributed 4 concept elements to this direction.
Computing Preferences Based on Agents' Beliefs (2014) — contributed 3 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_subjective_argument_evaluation_94f2f85c,
  title = {subjective argument evaluation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subjective-argument-evaluation/rdr_94f2f85c6b0b411bb1578ada997c93a1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_subjective_argument_evaluation_94f2f85c_v1,
  title = {subjective argument evaluation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/subjective-argument-evaluation/rdr_94f2f85c6b0b411bb1578ada997c93a1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake