Modeling complex narrative memory via sentence recognition: A marriage of distributional semantic models and an episodic model (MINERVA 2).

Recent research has integrated representations from distributional semantic models (DSMs) with traditional episodic memory models. While these integrated frameworks show promise in modeling single-word recognition, their applicability to the recognition of full sentences-particularly those embedded in complex narrative contexts-remains tentative. In this study, participants read short narratives and completed an OLD/NEW sentence recognition task, administered either immediately or after a delay of >10 hours. The recognition test contained four types of sentence probes: verbatim (matching the narrative word-for-word), paraphrase (conveying the same information with different wording), inference (logically inferred from the narrative but unstated), and wrong (information inconsistent with the narrative). We replicated prior findings that verbatim probes were judged OLD most often, followed by paraphrase, then by inference, and then by wrong probes. To model these behavioral findings, we imported semantic word representations from two DSMs, Latent Semantic Analysis and the Random Permutation Model, into MINERVA 2. Notably, we preregistered our modeling procedure and parameters for Study 2-one of the first in computational memory research-allowing us to generate predictions rather than postdictions. Overall, the MINERVA 2 model with DSM semantic representations performed very well and captured the key behavioral findings. Importantly, the integrated model substantially outperformed a version of the model where words were represented via randomly generated vectors, and also outperformed the DSM vectors when used on their own. These findings suggest that integrating semantic and episodic models of memory is crucial for accounting for the complexities of recognition of words, sentences, and whole narratives.