Our ability to ascertain which person a pronoun refers to is a central part of human language understanding. Toward a process-based understanding of the brain’s pronoun-resolution abilities, we evaluated four computational models against brain activity during naturalistic comprehension. These models each formalizes a different strand of explanation for pronoun resolution that has figured in the cognitive and linguistic literature. These include syntactic binding constraints, discourse coherence and principles of memory retrieval. We also examined a deep neural network model that has shown high performance in Natural Language Processing. We collected both functional Magnetic Resonance Imaging (fMRI) and magnetoencephalography (MEG) data while English and Chinese speakers listened to an extended narrative in the scanner. We applied univariate and multivariate analyses to correlate model predictions with brain activity patterns time-locked at each third person pronoun in the narratives. Our combined results all favor the memory-based model, suggesting a domain-general mechanism for pronoun resolution that resembles memory retrieval.
bioRxiv Subject Collection: Neuroscience