A Cognitively Grounded Bayesian Framework for Misinformation Susceptibility
This academic paper introduces the Bounded Pragmatic Listener (BPL), a novel cognitively grounded Bayesian framework designed to model human susceptibility to information disorder. Building upon Rational Speech Act theory, BPL incorporates three specific bounds derived from bounded rationality literature: a recursion depth bound reflecting working memory limits, a prior compression parameter addressing information bottlenecks, and an availability sample size utilizing saliency-weighted proposals. The framework aims to predict misinformation susceptibility, explain annotator disagreement, and differentiate vulnerability to mis-, dis-, and mal-information as defined in the Information Disorder framework. The authors validate the BPL model using the LIAR and MultiFC benchmarks, demonstrating competitive performance in veracity classification. Furthermore, the study provides experimental support for the depth-mismatch paradox. As a work in progress submitted to arXiv under Computer Science categories, this research contributes to the fields of computational linguistics and artificial intelligence by offering a robust theoretical and empirical approach to understanding how cognitive constraints influence the processing and acceptance of false or misleading information.
Wire timeline
A Cognitively Grounded Bayesian Framework for Misinformation Susceptibility
This academic paper introduces the Bounded Pragmatic Listener (BPL), a novel cognitively grounded Bayesian framework designed to model human susceptibility to information disorder. Building upon Rational Speech Act theory, BPL incorporates three specific bounds derived from bounded rationality literature: a recursion depth bound reflecting working memory limits, a prior compression parameter addressing information bottlenecks, and an availability sample size utilizing saliency-weighted proposals. The framework aims to predict misinformation susceptibility, explain annotator disagreement, and differentiate vulnerability to mis-, dis-, and mal-information as defined in the Information Disorder framework. The authors validate the BPL model using the LIAR and MultiFC benchmarks, demonstrating competitive performance in veracity classification. Furthermore, the study provides experimental support for the depth-mismatch paradox. As a work in progress submitted to arXiv under Computer Science categories, this research contributes to the fields of computational linguistics and artificial intelligence by offering a robust theoretical and empirical approach to understanding how cognitive constraints influence the processing and acceptance of false or misleading information.
cs.AI updates on arXiv.org