Question-answer dynamics and confirmation theory in reasoning by representativeness

Mathias Sablé-Meyer, Janek Guerrini, Salvador Mascarenhas


引用 0|浏览0
We show that probabilistic decision-making behavior characteristic of reasoning by representativeness or typicality arises in minimalistic settings lacking many of the features previously thought to be necessary conditions for the phenomenon. Specifically, we develop a version of a classical experiment by Kahneman and Tversky (1973) on base-rate neglect, where participants have full access to the probabilistic distribution, conveyed entirely visually and without reliance on familiar stereotypes, rich descriptions, or individuating information. We argue that the notion of evidential support as studied in (Bayesian) confirmation theory offers a good account of our experimental findings, as has been proposed for related data points from the representativeness literature. In a nutshell, when faced with competing alternatives to choose from, humans are sometimes less interested in picking the option with the highest probability of being true (posterior probability), and instead choose the option best supported by available evidence. We point out that this theoretical avenue is descriptively powerful, but has an as-yet unclear explanatory dimension. Building on approaches to reasoning from linguistic semantics, we propose that the chief trigger of confirmation-theoretic mechanisms in deliberate reasoning is a linguistically-motivated tendency to interpret certain experimental setups as intrinsically contrastive, in a way best cashed out by modern linguistic semantic theories of questions. These questions generate pragmatic pressures for interpreting surrounding information as having been meant to help answer the question, which will naturally give rise to confirmation-theoretic effects, very plausibly as a byproduct of iterated Bayesian update as proposed by modern Bayesian theories of relevance-based reasoning in pragmatics. Our experiment provides preliminary but tantalizing evidence in favor of this hypothesis, as participants displayed significantly more confirmation-theoretic behavior in a condition that highlighted the question-like, contrastive nature of the task.
AI 理解论文
Chat Paper