1 Title: Semantic associates create retroactive interference on an independent spatial recall task Authors: James W. Antony1, Kelly A. Bennion2 1Center for Neuroscience, University of California, Davis, Davis, CA, 95618, USA 2Department of Psychology and Child Development, California Polytechnic State University, San Luis Obispo, San Luis Obispo, CA Corresponding author: Dr. James Antony University of California, Davis [email protected] Center for Neuroscience Phone: (262) 347-8224 Davis, CA 95618, USA Funding: This work was supported by the Princeton University CV Starr Fellowship to JWA. Author note: Code / data for this project can be found here: Antony, J.W. & Bennion, K.A. (2020). Memory attack by semantic mediator: causing retroactive interference via Deese-Roediger-McDermott lists. Retrieved from osf.io/khrmx. 2 Abstract Semantic similarity between stimuli often leads to false memories, but how it causes retroactive interference (RI) in veridical recall has been less explored. Here, in Phase 1, participants learned spatial locations for “critical” words that reliably produce false memories in the Deese-Roediger-McDermott paradigm. In Phase 2, participants centrally viewed words that were semantically associated with half of the critical words. In Phase 3, participants retrieved the Phase 1 critical word locations. We found spatial memory RI for critical words whose semantic associates were shown (vs. not shown), suggesting that semantic relatedness caused retroactive interference. This effect was present in three experiments when the interfering information was presented shortly before spatial recall, but not after a one-hour delay between associate learning and test or after reversing the order of the critical word spatial learning phase and the associate learning phase. These findings suggest that memory impairments from RI can occur solely via semantic associates on an independent test where all relevant responses are freely available. We consider these findings to be an example of cue overload theory, whereby overload can interfere indirectly via semantic associates, and explain them by citing computational models of RI such as the search of associative memory theory. Keywords: Retroactive interference; episodic memory; semantic memory; false memory; cue overload 3 Long-term memory relies on encoding, retrieval, and processes (including new learning) that occur between encoding and retrieval. Retroactive interference (RI), or the process by which newly learned information interferes with previously learned information, has been shown for over a century to have a profound and reliable effect on memory (Müller & Pilzecker, 1900; Wixted, 2004). In one of the most common demonstrations of RI, participants forget more from a list of paired associates (A-B) when it is followed by a new list involving the same cue words (A-C) than when it does not (D-E). Interestingly, in a phenomenon that has been studied little since the late 1960s, RI also occurs when newly learned information is semantically similar, but not identical, to previously learned information. That is, memory for A-B pairs can be impaired by A’-C learning, where A and A’ are synonyms, relative to D-E learning (Baddeley & Dale, 1966; McGeoch & McDonald, 1931; Saltz & Hamilton, 1967). Generally, psychological theories related to these RI effects propose that new learning either causes the unlearning of previously learned information (Briggs, 1954) or acts independently of previously learned information and causes interference due to enhanced competition between previously and newly learned responses (Bower, Thompson-Schill, & Tulving, 1994; Mensink & Raaijmakers, 1988; Saltz & Hamilton, 1967; Slamecka & Ceraso, 1960). One form of competition is cue overload, or the finding that as more information becomes associated with a cue, any specific instance becomes less likely to come to mind at retrieval (Nairne, 2002; Robin, Garzon, & Moscovitch, 2019; Watkins & Watkins, 1976; Watkins & Watkins, 1975). However, it is unclear to what extent RI occurs via cue overload when information related to a cue, rather than the cue itself, is subsequently learned, and to what extent RI relies on the availability of responses at test. To address these questions in the present work, we investigated RI by presenting semantically similar, but not identical, information during new learning. Critically, this information did not interfere with the primary target response, as participants learned peripheral word locations on a spatial memory 4 task, but the interfering associations were all presented centrally. Moreover, all peripheral target locations were available at test, indicating that forgetting could not be attributed to the availability of the correct response. This contrasts with prior work in two ways. First, in most prior work, newly learned and previously learned information was identical (e.g., A cues), which, because participants often consciously notice this identical information, they can form strategies to overcome interference (e.g., Martin, 1968; Negley, Kelley, & Jacoby, 2018; Osgood, 1949). Second, in the studies employing semantic associates in long-term memory interference (e.g., Baddeley & Dale, 1966; McGeoch & McDonald, 1931; Saltz & Hamilton, 1967), the target responses (e.g., B cues) overlapped in their lexical nature with the new information (i.e., they were both words) and were also not available as choices at test. We created semantic similarity between previously and newly learned information by using lists of words from the Deese-Roediger-McDermott (DRM) paradigm (Deese, 1959b; Roediger & McDermott, 1995). In this paradigm, lists of words (e.g., marker, crayon) reliably elicit false recall and recognition for one unpresented critical word to which they are all related (e.g., pen). Notably, this false memory effect seems to rely partly on the backward associative strength (BAS) between the words (Deese, 1959a; Roediger, Watson, McDermott, & Gallo, 2001), which is a measure of how often critical lure words are freely given when prompted by the associates (e.g., marker pen) in free association tests (Nelson, McEvoy, & Schreiber, 1998). Here, rather than assessing false memory for critical words, we used the related associates to create RI after presenting the critical words and measured RI as a function of BAS. Figure 1 shows the procedures within and the sequencing of the three phases used in the five experiments. In the spatial encoding phase of Experiment 1, participants learned spatial locations for critical words (e.g., pen) (A-B learning, with critical words indicated by A and locations by B) (Fig 1). In the associate learning phase, participants centrally viewed words that were semantically related associates of half of the critical words (e.g., marker, crayon) (A’-C learning, with related associates 5 indicated by A’ and the center location as C). To reduce the likelihood of participants being able to notice the similarity between these associates and the critical words (and thereby enabling the ability to form conscious strategies), these lists of associates were intermixed with each other and shown in a random order. Finally, participants retrieved the spatial locations for the critical words (to test recall for A-B) and distinguished previously seen critical words from novel words (to test recognition for A). These manipulations allowed us to determine whether interference via cue overload could occur in the absence of 1) direct cue competition, as the original (critical words) and interpolated material (associated words) differed; 2) direct target competition, as the original (peripheral locations) and interpolated material (central locations) differed; 3) interference due to response availability, as all spatial responses were available at test; and 4) participants’ awareness of whether the interpolated words were related to the original cue words. To foreshadow, we found that critical word spatial memory, but not recognition, was significantly worse for words whose related associates were presented (vs. not presented) in the associate learning phase, an effect we pre-registered and directly replicated in Experiment 2. In Experiment 3, we swapped the order of these two phases and found no spatial memory RI for words whose related associates were previously presented; instead, we found an improvement in recognizing critical words whose related associates were previously shown. In Experiments 4 and 5, we contrasted two possible accounts for the above findings: a reactivation/storage account and a retrieval impairment account. If presenting associates leads to relatively long-lasting RI (e.g., 60 minutes), spatial recall for the related condition (relative to the control condition) should remain impaired after a 60-minute delay between presentation of related associates and final test (Experiment 4). According to the reactivation/storage account, RI would occur if the memory trace for the critical word-location association (A-B) were reactivated to a modest extent during the subsequent associate phase (A’-C), which, according to behavioral and neural evidence, would result 6 in weakening or suppression of the previously stored long-term memory traces for A-B (Anderson, 2003; Lewis-Peacock & Norman, 2014; Norman, Newman, & Detre, 2007). However, if presenting associates shortly before testing leads to greater temporary retrieval interference for the critical word-location associations due to cue overload, as shown previously (Anderson