The Bayesian Echo Chamber: Modeling Social Influence via Linguistic Accommodation

Fangjian Guo, Charles Blundell, Hanna M. Wallach, Katherine A. HellerAISTATS 2015 · OpenReview

Abstract

We present the Bayesian Echo Chamber, a new Bayesian generative model for social interaction data. By modeling the evolution of people’s language usage over time, this model discovers latent influence networks from transcripts of group discussions. Unlike models based on turn-taking, our approach leverages linguistic accommodation—the subconscious tendency of individuals to adapt their language to those they interact with—to reveal deeper social hierarchies and power dynamics.

Introduction

In social science and network analysis, determining "who influences whom" is a fundamental challenge. Because influence is rarely stated explicitly, researchers often rely on proxies. Traditionally, this meant looking at structural links or temporal dynamics, such as turn-taking behavior (e.g., if Alice speaks, does Bob usually speak next?).

However, turn-taking is a noisy signal. A person might speak often simply because they are talkative or because they are being interrogated, not because they are influential. We argue that content—specifically how word choices drift to match those of interlocutors—is a much more potent signal of influence. This phenomenon, known as linguistic accommodation, suggests that less powerful individuals tend to adapt their language to resemble more powerful ones.

The Bayesian Echo Chamber (BEC) Model

The authors introduce the Bayesian Echo Chamber (BEC), a dynamic language model that combines Bayesian language modeling with a discrete analog of the multivariate Hawkes process.

  • The Mechanism: Every participant has an "inherent" language usage profile. During interaction, the use of a word by one person increases the probability of another person using that same word shortly thereafter.
  • Latent Influence: The model includes latent variables ($\rho^{(qp)}$) representing the degree of linguistic excitation from person $q$ to person $p$. If person $p$ frequently "echoes" person $q$, the model infers that $q$ has high influence over $p$.
  • Temporal Decay: Similar to a Hawkes process, the influence of an utterance decays over time, meaning the "echo" effect is strongest immediately following an interaction.
  • Inference: A collapsed slice-within-Gibbs sampling algorithm is used to infer latent influence networks and language parameters from raw transcripts.

Datasets

The model was evaluated on three distinct types of social interaction data:

  1. US Supreme Court Transcripts: Three controversial cases (Lawrence v. Texas, DC v. Heller, and Citizens United).
  2. "12 Angry Men" (1957): A movie transcript of a jury deliberation, providing a naturalistic setting for consensus-building.
  3. Federal Open Market Committee (FOMC) Meetings: Transcripts from 32 meetings surrounding the 2007–2008 financial crisis, representing complex group decision-making.

Key Insights and Ideas

Talkative vs. Influential

Turn-taking models often mistake "interrogated" or "verbose" people for influential ones. In Supreme Court data, turn-taking models suggest Justices influence attorneys because Justices interrupt to ask questions. The BEC model correctly identifies that the attorneys exert influence by setting the linguistic agenda that the Justices then adopt.

Exerted vs. received influence in the Supreme Court case DC v. Heller. Attorneys (DELLI, GURA, CLEME) show high exerted influence compared to Justices.

Context-Dependent Influence

In the FOMC data, influence shifted based on the economic climate. During "tightening" periods (pre-crash), "hawks" were highly influential. As the crisis hit and policy shifted to "easing," their influence diminished in favor of "doves" and close allies of Chairman Bernanke.

Key Results

Predictive Performance and Sparsity

The Bayesian Echo Chamber consistently outperformed standard unigram models and Dynamic Topic Models (DTM) in terms of the log-probability of held-out data across all datasets. Furthermore, the influence networks produced by BEC are generally sparser and more substantively meaningful than those produced by turn-taking models, which often result in dense, uninterpretable "hairballs" of connections.

Latent influence network for DC v. Heller. The model discovers a sparse, interpretable structure of social influence.

Joint Modeling

When the authors combined turn-taking and linguistic accommodation into a single model, they found that the linguistic data was the dominant and more informative signal for discovering global influence.

Conclusion

The Bayesian Echo Chamber provides a robust framework for uncovering latent social dynamics through the lens of linguistic accommodation. By moving beyond simple turn-taking, it offers a more nuanced understanding of power and influence in group settings. Future work may involve distinguishing between function and content words or modeling continuously evolving influence networks.

Summary updated

Abstract

We present the Bayesian Echo Chamber, a new Bayesian generative model for social interaction data. By modeling the evolution of people’s language usage over time, this model discovers latent influe...
BibTeX
@inproceedings{guo2015bayesian,
  author = {Fangjian Guo and Charles Blundell and Hanna M. Wallach and Katherine A. Heller},
  title = {{The Bayesian Echo Chamber: Modeling Social Influence via Linguistic Accommodation}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2015},
  url = {https://openreview.net/forum?id=K-0veDt_y8},
}
Figure