Fig. 5: Learning-induced changes in representational similarity for paired words.
A Linear mixed effects modeling revealed that pairs of words that were experienced as to-be-learned pairs (regardless of whether tested or restudied) and that were correctly recalled at final test (regardless of learning manipulation; orange bar) became more similar after learning than pairs that were not successfully recalled at final test (pink bar; t(158) = 2.566, p = 0.023, d = 0.20, 95% CI = [1.84 × 10−5 6.29 × 10−4]) and arbitrary pairings of words that were never experienced as to-be-learned pairs (blue bar; t(158) = 3.112, p = 0.007, d = 0.25, 95% CI = [8.73×10−5 6.98 × 10−4]). Y-axis indicates change in similarity (post-learning assessment minus pre-learning assessment). B Two-tailed one-sample t-tests reveal significant learning-induced similarity change for pairs of words that were tested (regardless of semantic relatedness; related: t(79) = 3.788, p = 0.002, d = 0.423, 95% CI = [3.045 × 10−4 9.804 × 10−4]; unrelated: t(74) = 3.085, p = 0.017, d = 0.356, 95% CI = [2.736 × 10−4 1.272 × 10−3]) and for semantically related restudied pairs (t(79) = 4.258, p < 0.001, d = 0.476, 95% CI = [3.763 × 10−4 1.037 × 10−3]). Green indicates semantically related pairs, yellow indicates semantically unrelated pairs. For both panels A and B, open circles reflect means of individual participants (N = 80), with connecting lines showing within-subject differences across conditions. C Words that may interfere with successful recall of to-be-learned pairs (i.e. lures) were defined by their LSA cosine similarity to a given to-be-learned cue. Change in similarity of these potential lure pairs was calculated across learning and entered into a linear mixed effects model based on data from 248,408 pairs of words across 80 participants. This model revealed a significant learning condition by lure strength interaction (t(248400) = 2.840, p = 0.005, ηp2 = 3.25 × 10−6, 95% CI = [−3.130 × 10−4 −5.740 × 10−5]), where very strong lures are drawn towards a given cue word regardless of learning condition (restudy: z = 6.689, p < 0.001, d = 0.15, 95% CI = [6.59 × 10−4 1.48 × 10−3]; test: z = 6.029, p < 0.001, d = 0.12, 95% CI = [4.90 × 10−4 1.22 × 10−3]), while strong lures showed no statistically significant change relative to weak/non-lures (restudy: z = 1.707, p = 0.32, d = 0.022, 95% CI = [−3.94 × 10−4 7.93 × 10−5]; test: z = 1.964, p = 0.200, d = 0.022, 95% CI = [−3.76 × 10−4 5.02 × 10−5]). In contrast, moderate lures are pushed away from cue words (relative to baseline weak/non-lures) more when the associated to-be-learned pairs are tested than restudied (z = 2.840, p = 0.014, d = 0.03, 95% CI = [5.740 × 10−5 3.129 × 10−4]). Maroon lines reflect lures associated with restudied pairs, blue lines reflect lures associated with tested pairs. Dotted line reflects average change in similarity for baseline weak/non-lure pairs. Across all panels, error bars reflect standard error of the mean. Symbols reflect statistically significant differences using two-tailed tests across conditions (panels A and C) or versus zero (panel B) using Holm-Bonferroni corrections for multiple comparisons (*p < 0.05, **p < 0.01, ***p < 0.001).
