Memorability of novel words correlates with anterior fusiform activity during reading

Abstract

Our memory for the words we already know is best predicted by their associated meanings. However, the factors that influence whether we will remember a new word after we see it for the first time are unclear. We record memory performance for 2100 novel pseudowords across 1804 participants during a continuous recognition task. Participants show significant agreement across individuals for which novel words were memorable or forgettable, suggesting an intrinsic memorability for individual pseudowords. Pseudowords that are similar to low-frequency known words, with sparse orthographic neighbourhoods and rarely occurring letter pairs, are more memorable. Further, using intracranial recordings in 36 epilepsy patients we show a region in the anterior fusiform cortex that shows sensitivity to the memorability of these pseudowords. These results suggest that known words in our lexicon act as a scaffold for remembering novel word forms, with rare and unique known words providing the best support for novel word learning.

Fig. 1. Experimental design and behavioural analysis.

a Schematic representation of the continuous recognition task. b Distributions of participant hit rates, false alarm rates and response times during the continuous recognition task. Box plots represent quartiles, with whiskers extending to ±1.5 × IQR. Density plots were calculated using the random average shifted histogram method. c Spearman-Brown split-half analysis of the memorability results, testing 1000 randomly split halves of the data against shuffled data.

Fig. 2. Quantification of novel word memorability.

a Histogram of pseudoword memorabilities within our 2100-item corpus. b Exemplar items showing the highest and lowest memorability items for each word length. c Predictability of item memorability based on multiple linear regression, showing the Spearman rank correlation coefficient with 95% confidence interval.

Fig. 3. Distribution of novel word memorabilities.

Spatial map of the pseudoword corpus with the vertical axis representing memorability and horizontal axis distributed by word length. Words are coloured by orthographic neighbourhood, with sparser neighbourhoods being darker.

Fig. 4. Neighbourhood density of memorable pseudowords.

Difference (mean ± SE) in neighbourhood size at varying Levenshtein (character-edit) distances for pseudowords of each word length. Low (<20th percentile; blue) and high (>80th percentile; red) memorability score pseudowords’ neighbourhood sizes were calculated as their percentage difference from median memorability (40–60th percentile) pseudowords within each word length. Low and high memorability neighbourhood sizes were compared using two-tailed Wilcoxon rank-sum tests; *p < 0.05, **p < 0.01, ***p < 0.001. Number of pseudowords per quintile is shown for each word length.

Fig. 5. Intracranial correlates of novel word memorability.

a, b sbLME models for a memorability and b lexicality, showing significant clusters (p < 0.01, Monte Carlo cluster corrected) of sensitivity. Regions in black did not have sufficient coverage for reliable sbLME results (<3 patients). c Anatomical ROI definitions for posterior (pFus; dark green), mid (mFus; green), and anterior (aFus; light green) fusiform cortex. Pseudoword responsive electrodes not included in an ROI shown in white. d LME beta values (±SE) for the effect of memorability within each ROI, calculated at each time point independently (10 ms resolution). Effects of word position in the sequence, orthographic neighbourhood, bigram and quadrigram frequencies, and lowest frequency neighbour were regressed out. Number of electrodes and patients per cluster is shown. Coloured bars represent regions of significance from the LME analyses (q < 0.01 FDR-corrected).

Fig. 6. Novel word memorability outside of vOTC.

a Anatomical ROI definitions for inferior frontal gyrus (IFG; dark purple), precentral sulcus (pCS; purple), and inferior parietal sulcus (IPS; light purple). Pseudoword responsive electrodes not included in an ROI shown in white. b LME beta values (±SE) for the effect of memorability within each ROI, calculated at each time point independently (10 ms resolution). Effects of word position in the sequence, orthographic neighbourhood, bigram and quadrigram frequencies, and lowest frequency neighbour were regressed out. Number of electrodes and patients per cluster is shown. None of these ROIs showed any periods of significant sensitivity to memorability (all q > 0.05 FDR-corrected).