Scope interpretation in natural and artificial language processing

This study examines the processing and interpretation of quantified sentences in the a-every configuration across English and Chinese using self-paced reading, comprehension questions, and large language model (LLM) comparisons. We addressed three research objectives: (1) to determine whether inverse scope interpretations incur greater processing costs than surface scope and whether these costs differ between English and Chinese, (2) to investigate how working memory (WM) influences real-time processing and offline interpretation of quantifier scope, and (3) to evaluate the extent to which surprisal-based measures from autoregressive models, including GPT-2 small (English and Chinese variants), BLOOM-560M, and Qwen2.5-0.5B, align with human processing patterns. Results revealed longer reading times for inverse scope, where inverse scope interpretations were overall less accessible than surface scope in both English and Chinese. Offline comprehension questions indicated a much strong preference for surface scope in Chinese than in English, with WM modulating offline but not real-time processing, suggesting distinct processes involved real-time processing and offline interpretation. Surprisal analyses showed that the LLMs captured the global patterns of scope interpretation, particularly in Chinese, as reflected in whole-sentence surprisal. The strongest correspondence with human data, however, was observed for Qwen2.5-0.5B in English, and overall alignment with human processing varied by model size and training data. Overall, the findings highlight cross-linguistic asymmetries in scope accessibility, differential cognitive demands underlying real-time and offline interpretation, and both the potential and current limitations of LLMs in modeling and predicting human sentence processing.