Density plots show the distribution of within-participant correlations between behavioural representational dissimilarity matrices (RDMs) and RDMs obtained from phonological and semantic representations of the 18 target words (medians, interquartile ranges and densities). Behavioural RDMs were obtained by computing the confusion matrix from participants’ behavioural responses. Pairwise phonological distances for the stimulus material were computed using Phonological Corpus Tools (V1.4.0), and semantic distances were computed using fastText vector representations (see Materials and methods for details). Representational similarity was computed using Spearman rank correlations. Overall, phonological and semantic features both influenced participants’ responses. A repeated-measurements ANOVA (2 (conditions) x 2 (features)) yielded a main effect of condition (F(1,19) = 7.53, p=0.013; mean correlations: Mauditory = 0.38, SEM = 0.01; Mvisual = 0.43, SEM = 0.02) and a main effect of features (F(1,19) = 20.98, p = <0.001, Mphon = 0.43, SEM = 0.01; Msem = 0.37, SEM = 0.01). A post-hoc comparison revealed that in both conditions phonological features influenced behaviour stronger than semantic features (Wilcoxon Signed-ranks test; Zauditory = 151, p=0.037, Zvisual = 189, p=0.002). Examples for close phonological and semantic relationships between two words are given in image.