Accumulating evidence for myriad alternatives: Modeling the generation of free association.

Abstract

The associative manner by which thoughts follow one another has intrigued scholars for decades. The process by which an association is generated in response to a cue can be explained by classic models of semantic processing through distinct computational mechanisms. Distributed attractor networks implement rich-get-richer dynamics and assume that stronger associations can be reached with fewer steps. Conversely, spreading activation models assume that a cue distributes its activation, in parallel, to all associations at a constant rate. Despite these models' huge influence, their intractability together with the unconstrained nature of free association have restricted their few previous uses to qualitative predictions. To test these computational mechanisms quantitatively, we conceptualize free association as the product of internal evidence accumulation and generate predictions concerning the speed and strength of people's associations. To this end, we first develop a novel approach to mapping the personalized space of words from which an individual chooses an association to a given cue. We then use state-of-the-art evidence accumulation models to demonstrate the function of rich-get-richer dynamics on the one hand and of stochasticity in the rate of spreading activation on the other hand, in preventing an exceedingly slow resolution of the competition among myriad potential associations. Furthermore, whereas our results uniformly indicate that stronger associations require less evidence, only in combination with rich-get-richer dynamics does this explain why weak associations are slow yet prevalent. We discuss implications for models of semantic processing and evidence accumulation and offer recommendations for practical applications and individual-differences research.

Complex, Dynamic, & Internal: As Simple As Possible, But No Simpler Than That

“…Regardless of the type of internal object (verbal thought, episodic memory), all objects are tied to selection by these nets. Indeed, exploring this assumption may have implications for integrating related theories of associative thinking (Andrews-Hanna et al, 2021;Fradkin & Eldar, 2022) together with A2T (in-line with Koster et al's (this issue) call for integrative and cumulative theory and science of IDC); and more broadly, raising questions and possible directions for future research into the effects of different internal objects and processes on the dynamics of IDC.…”

“…Indeed, A2T rests on an implicit assumption (see Model Components > Internal attentional selection section in the target article) that representations in WM send "activations" along associative semantic nets (e.g. Fradkin & Eldar, 2022). This is how different objects become activated in A2T and compete for selection.…”

Lexical search and social reasoning jointly explain communication in associative reference games

“…In the present work, we derived associative strengths from a large database of free association norms (De Deyne et al, 2019). However, free association itself is a process (Jones et al, 2015), and precisely how free associations are generated, i.e., what comes to mind, is an active area of research (Fradkin & Eldar, 2022;Richie et al, 2022;Kumar, 2021). Ultimately, a complete account of lexical search and selection will require an explanation for the process of generating free associations, that converges with how these processes interact with those of search and selection explored in the current work.…”

“…Olson, 1970;Dale & Reiter, 1995;Degen, Hawkins, Graf, Kreiss, & Goodman, 2020) but are restricted to narrow lexical domains. Meanwhile, models of semantic search are typically studied through free-association tasks (Fradkin & Eldar, 2022;Richie et al, 2022;Kumar, 2021) or fluency tasks (Rhodes & Turvey, 2007;Zemla & Austerweil, 2018;Hills et al, 2015) without an explicit communicative context. Understanding how internal lexical processes are integrated with pragmatic selection processes thus requires an appropriately rich and open-ended paradigm that draws on both processes.…”

The effect of memory load on object reconstruction: Insights from an online mouse-tracking task

“…However, while 20 trials per set size was sufficient for our specific purposes, future work will need to consider whether other research questions will require more trials. Approaches involving model-fitting (e.g., for models applied to working memory, see Bays, 2016, Oberauer, 2022, Rademaker et al, 2018Schurgin, Wixted, & Brady, 2020;Sutterer & Awh, 2016;Zhang & Luck, 2008; also see evidence accumulation models: Bornstein et al, 2018;Evans & Wagenmakers, 2020;Fradkin & Eldar, 2022;Krueger et al, 2017;Shenhav, Straccia, Musslick, Cohen, & Botvinick, 2018) or individual differences (e.g., Baker et al, 2020;Xu et al, 2017) will have different considerations.…”