Semantic relatedness retroactively boosts memory and promotes memory interdependence across episodes

Abstract

Two fundamental issues in memory research concern when later experiences strengthen or weaken initial memories and when the two memories become linked or remain independent. A promising candidate for explaining these issues is semantic relatedness. Here, across five paired-associate learning experiments (N=1000), we systematically varied the semantic relatedness between initial and later cues, initial and later targets, or both. We found that learning retroactively benefited long-term memory performance for semantically related words (vs. unshown control words), and these benefits increased as a function of relatedness. Critically, memory dependence between initial and later pairs also increased with relatedness, suggesting that pre-existing semantic relationships promote interdependence for memories formed across episodes. We also found that modest retroactive benefits, but not interdependencies, emerged when subjects learned via studying rather than practice testing. These findings demonstrate that semantic relatedness during new learning retroactively strengthens old associations while scaffolding new ones into well-fortified memory traces.

Keywords: human; memory integration; memory reactivation; neuroscience; retroactive facilitation; retroactive interference; semantic memory.

Figure 1.. Overview of conditions, stimuli, experiments, variables, and Osgood’s predictions.

(a) After base pair learning, pairs were divided into five experimental conditions for secondary pair learning. After a 5-min or 48-hr delay, both base and secondary pairs were tested. (b) Word pairs were counterbalanced every five subjects into conditions. (c) Overview of the first four experiments by delay and stimulus set. (d) Coverage of variables across associative strength (AS) values in the stimulus set with a narrower range of semantic relationships. Crosses along the ΔCue (green) and ΔTarget (blue) lines show distributions of cue and target relatedness, respectively. Purple crosses inside the surface (scatterplot) show the distribution of bivariate cue and target relatedness in the ΔBoth condition. (e) Experimental data hypothetically conforming to Osgood, 1949 proposed surface. Cue and target relatedness span the y- and x-axes, respectively, while memory change for each condition spans the z-axis, relative to the control condition on the z=0 surface. In (d) and (e), example word pairs from (a) were labeled for illustrative purposes. The x-axes were reversed from normal convention to correspond to Osgood’s surface. See also Figure 1—figure supplement 1 for visualizations using the stimulus set with wider semantic relatedness and Supplementary files 1-2 for all stimuli.

Figure 1—figure supplement 1.. Wider stimulus set examples and coverage.

(a) Sample stimuli from the wider stimulus set in all conditions. (b) Variable coverage. Points along the target identity (green) and cue identity (blue) lines show distributions of cue and target GloVe values, respectively. Purple points inside the surface (scatterplot) show the distribution of bivariate GloVe values in the ΔBoth condition.

Figure 2.. RF versus RI differed by overall stimulus set relatedness, delay, and word pair condition.

The narrower stimulus set (top row) featured only single-step semantic associations between base and secondary cues and targets, whereas the wider stimulus set (bottom row) featured a full range of semantic relationships. All comparisons were significant except those labeled with gray bars and ‘ns’ (p>0.1) or † (0.05

Figure 2—figure supplement 1.. Secondary pair memory changes depending on overall stimulus set relatedness, delay, and word pair condition.

All comparisons were significant except those labeled with gray bars and ‘ns’ (p>0.1) or † (0.05

Figure 3.. Target relatedness retroactively benefited memory and created interdependence between base and secondary pairs.

(a) Across-subject memorability for each base pair was plotted against the target semantic relatedness, with AS and GloVe values in the top and bottom rows, respectively. Relatedness improved memory in all experiments except in the narrower stimulus set, 5-min delay experiment, where overall memory approached ceiling performance. RI occurred with especially low relatedness in the wider stimulus set, 5-min delay experiment, but this interference disappeared with high relatedness. In the wider stimulus set, 48-hr delay experiment, we found no interference with low relatedness and facilitation with high relatedness. (b) Within each base pair target-secondary pair target duo, we plotted across-subject memory dependence against semantic relatedness in all experiments. Higher correlation values indicate that subjects tended to remember or forget both targets in the duo together. Relatedness increased these correlations in three experiments, excluding the narrower stimulus set, 5-min delay experiment. Thick dotted lines show the 95th percentile threshold of dependence levels against all other pairs. See also Figure 3—figure supplement 1 for intrusion data from this condition. RF, retroactive facilitation; RI, retroactive interference.

Figure 3—figure supplement 1.. Intrusions did not increase with target relatedness in the ΔTarget condition.

(a) Across-subject intrusion rates of secondary pairs into each base pair were plotted against target relatedness (Top: AS; bottom: GloVe). We found no positive relationships between intrusions and target relatedness, with the wider stimulus set, 5-min delay experiment featuring a significantly negative relationship.

Figure 4.. Cue semantic relatedness has no consistent retroactive effect.

Across-subject memorability for each base pair – control was plotted against cue relatedness (top: AS; bottom: GloVe). Relatedness had no effect on memory in any condition. See also Figure 4—figure supplement 1 for dependence correlations.

Figure 4—figure supplement 1.. Cue relatedness showed a mixed relationship with memory dependence in the ΔCue condition.

Within each base pair target-secondary pair target duo, we plotted across-subject memory dependence against relatedness in all experiments. Relatedness increased these correlations in the narrower stimulus set, 5-min delay and the wider stimulus set, 48-hr delay experiments. Thick dotted lines depict the 95th percentile threshold of dependence levels against all other pairs.

Figure 5.. High cue and target relatedness promoted long-term RF and increased interdependence.

(a) We plotted a smoothed surface of ΔBoth – control memorability values against cue and target relatedness on the x- and y-axes (top: AS; bottom: GloVe). Under the narrower stimulus set in the 48-hr delay experiment, memorability increased at high levels of both cue and target relatedness. (b) Similar to (a) except depicting smoothed surfaces of base-secondary pair dependence. High values of cue and target relatedness increased dependence in the narrower stimulus set, 48-hr delay experiment. Sections of these plots with purple grid marks were significant, whereas those with white marks were not. See Figure 5—figure supplement 1 for linear contrasts between cue+target relatedness and memory and dependence measures. RF, retroactive facilitation.

Figure 5—figure supplement 1.. High cue+target relatedness promotes long-term RF and increases interdependence.

(a) We plotted across-subject memorability for each base pair – control against the added cue+target relatedness (top: AS; bottom: GloVe). Under the narrower stimulus set in the 48-hr delay experiment, memorability increased with relatedness. (b) We plotted across-subject dependence for each base pair target-secondary pair target duo against added relatedness in all experiments. Relatedness increased dependence in the narrower stimulus set, 48-hr delay experiment. Thick dotted lines show the average 95th percentile threshold of dependence levels against all other pairs. RF, retroactive facilitation.

Figure 6.. Osgood-style surfaces depicting retroactive effects and dependence.

(a) We plotted all conditions (vs. control) from all experiments in three-dimensional coordinates, with cue and target relatedness on the y- and x-axes, respectively, and retroactive memory change on the z-axis, with RF and RI in the positive and negative directions, respectively. ((a), right) For the narrower stimulus set, 48-hr delay experiment, we plotted memory for the No Δ – control condition (±across-pair standard deviation) at the cue identity, target identity corner point (red circle). We plotted ΔTarget – control condition memory along the cue identity line against target relatedness (± standard error from the ordinary-least-squares regression fit) (blue), and we plotted ΔCue – control condition memory along the target identity line against cue relatedness (± standard error from the ordinary-least-squares regression fit) (green). We plotted ΔBoth – control condition memory as a locally smoothed surface as a bivariate function of cue and target relatedness (purple). Transparent surface grids above and below zero represent p<0.01 significance boundaries from permutation tests, beyond which the surface is significant, as indicated by a darker shade of purple. (left) Similar plots created for all conditions from the other experiments. (b) Dependence for all experiments and conditions formatted similarly to (a). RF, retroactive facilitation; RI, retroactive interference.

Figure 6—figure supplement 1.. Semantic relatedness benefits memory and dependence when combining data sets under a common GloVe metric.

(a) We plotted across-subject memorability (top) and dependence (bottom) for the ΔTarget (left), ΔCue (middle), and ΔBoth (right) conditions after combining the narrower and wider stimulus set, 5-min delay experiments. For the ΔBoth condition, we used added cue+target relatedness. We found significant relationships between relatedness and both memorability and dependence in the ΔTarget and ΔCue conditions. Circles and crosses depict narrower and wider stimulus set data, respectively. Correlations between the combined correlation are plotted in dark colors, with each individual correlation between stimulus set and relatedness replotted in light colors for visual comparison. Text labels reflect correlation values for the combined data sets. (b) We created identical plots except combining the experiments with 48 hr delays. Here all relatedness correlations were significant. This figure is linked with Figure 6 but is related to Figures 3—6.

Figure 7.. Study-only learning led to semantic relatedness benefits in the ΔTarget condition but no base-secondary pair dependence.

(a) Overall memory performance for base (left) and secondary pairs (right) by condition. All comparisons were significant except those labeled with gray bars and ‘ns’ (p>0.1) or † (0.05<p<0.1). Data points from individual subjects were jittered slightly for better visualization. (b) Correlations between retroactive memory effects (left) and base-secondary pair dependence (right) in the ΔTarget (top), and ΔCue (middle) conditions. Retroactive memory effects correlated with target relatedness in the ΔTarget condition, but no other comparisons were significant. Pearson correlations are shown in the plots followed by * when p<0.05 and ** when p<0.01. On bottom, we plotted retroactive (left) and dependence surfaces (right) in the ΔBoth condition.

Figure 8.. Secondary pair learning differed by stimulus set and condition and generally benefitted from semantic relatedness.

(a) Learning time (mean trials to criterion) followed this pattern for the narrower stimulus set: No Δ ‘ns’ (p>0.1) or † (0.05<p< 0.1). Data points from individual subjects were jittered slightly for better visualization. (b) In the ΔCue condition, average learning time across subjects for each word pair decreased with increasing cue relatedness (top: AS; bottom: GloVe). (c) In the ΔTarget condition, learning time generally decreased with increasing B/B′ relatedness (top: AS; bottom: GloVe). One exception occurred for the narrower stimulus set, 48-hr delay experiment. (d) In the ΔBoth condition, learning time decreased with cue+target relatedness in the narrower stimulus set, but not in the wider stimulus set. In (b–d), Pearson correlations are shown in the plots followed by * when p<0.05 and ** when p<0.01.

Figure 9.. Proposed neurobiological schematic of long-term memory fates across conditions and two levels of semantic relatedness.

Initially, unrelated base pair associations are bound in the hippocampus along with base list (BL) contexts (e.g., ‘plaza-doze-base list’ and ‘beer-late-base list’; left column), followed by new associations in secondary list (SL) contexts. Subjects can thereby retrieve target words given a cue and list context. Cortical associations are absent for words without pre-experimental relatedness (top) and strong for words with high pre-experimental relatedness (bottom). Relearning A-B (No Δ condition; second from left column) under either level of relatedness results in strengthened base pair and robust secondary pair associations (e.g., ‘plaza-doze-secondary list’), likely as linked episodes within the hippocampus. The consequences for learning a pair with a new target (ΔTarget condition; middle column) differs based on pre-experimental relatedness: with no relatedness (top), little to no long-term change occurs to the base pairs and secondary pairs are independently formed (e.g., ‘plaza-plus-secondary list’), whereas with high relatedness (bottom), the original memory is strengthened as the new memory is formed (keg-late-secondary list). Moreover, the episodes become bound within the hippocampus, forming an interdependent memory trace. Note that RI effects shown in the narrower stimulus set, 5-min delay experiment are not represented by this long-term schematic and likely stem from more temporary retrieval impairments. Learning a pair with a new cue (ΔCue condition; second column from right) generally strengthens the original memory (perhaps due to better target accessibility) in a manner only weakly reliant on cue relatedness. The consequences for learning a pair with two new words (ΔBoth condition; rightmost column) differ markedly based on pre-experimental relatedness: with no relatedness (top), the new association memory is formed independently, whereas with high relatedness (bottom), the base pair memory becomes strengthened, and the base and secondary pair associations become interdependent. RI, retroactive interference.