Investigating Referential Adaptation:The Statistics and Learning Mechanism

Pronoun resolution is guided by various discourse cues. For instance, ambiguous pronouns are often assigned to the subject of the preceding sentence (e.g., Ana went shopping with Liz. She bought a hat). However, how these biases develop has not been widely discussed. Studies have found that people learn discourse biases through exposure, tracking frequent pronoun-reference patterns (e.g., pronouns often refer to subjects), a process known as referential adaptation (e.g., Kaiser, 2009; Johnson & Arnold, 2023; Ye & Arnold, 2023b, etc.). Yet, the specific patterns people learn and the mechanisms underlying this learning remain unclear. Two experiments investigated the specificity of referential adaptation by testing whether third-person pronouns (e.g., he vs. she) are categorized as a general class or individual lexical items and whether referents are categorized by their grammatical positions (e.g., subject vs. non-subject) across different event roles (goal/source vs. agent/comitative). Results indicate that referential adaptation occurs at the level of pronoun categories rather than specific lexical items, and that referents are grouped by their grammatical positions regardless of semantic variations. Another two experiments examined whether referential adaptation is driven by prediction error, which primarily accounts for learning of unexpected patterns, and minimal or no learning when there is no unexpected input. However, comparable referential adaptation occurred both in strongly-biased contexts with unexpected references and equi-biased contexts lacking expectation violations, challenging the necessity of prediction error for referential adaptation. Alternative frameworks are evaluated and discussed in relation to existing findings on referential adaptation, emphasizing the need for a comprehensive model that integrates different results currently explained separately by error-based and frequency-based learning accounts.