MINERVA OPS: A computational framework for the representation and recognition of orthographic, phonological, and semantic associates.

Remembering is constructive. Thus, a complete account of memory must predict patterns in both veridical and false recognition. Ideally, that account will predict the content of what people falsely remember, not just that they falsely remember. To move towards that goal, we need theories of both representation and retrieval as well as an account that joins those two accounts into a cohesive whole. To work towards that goal, Reid and Jamieson Journal of Memory and Language, 129, 104397, (2023) previously modelled false recognition in the Deese-Roediger-McDermott paradigm by importing vectors that represent word meaning, derived from distributional semantic models, into the MINERVA 2 framework. However, semantic meaning is only one aspect of a word's representation in human memory as words are also represented in terms of their perceptual features (i.e., their phonology and orthography). In this paper, we extended Reid and Jamieson's account by developing a common method to derive orthographic and phonological representations for words to investigate veridical and false recognition of words in DRM-like tasks. To test the model, we conducted three false recognition experiments for study lists composed of (a) semantic word associates, (b) phonological neighbors, and (c) perceptually similar nonwords. Following this, we conducted two more experiments where phonological and orthographic similarity were more delineated, with critical lures that were either orthographically, but not phonologically related to studied items (Exp. 2a) or vice versa (Exp. 2b). We then extend the analysis to three more experiments with the directed forgetting procedure. The model tracks people's recognition and false recognition of lures that are semantically, phonologically, and orthographically related to studied items. These findings connect current efforts back to Herbert Simon's arguments about the importance of modelling both representation and process for understanding memory. It also joins a growing body of work on the development of structured language representations and the incorporation of those representations into existing models of memory. At broader view, we conclude that the field's models of memory are not only positioned to explain people's memory performance in principle, but also positioned to investigate and predict people's memory performance conditional on the specific memoranda presented in experimental tasks.