A distinctive meaning makes a sentence memorable

Abstract

Prior work on visual memory has suggested that humans have a high-capacity but imperfect memory: image representations accumulate noise over time, which makes similar images confusable. This account – the noisy representation hypothesis – was recently extended to the verbal domain: in line with past evidence that words are encoded in memory by their meanings, it was shown that words with distinctive meanings are most memorable. Here, we leverage recent advances in natural language processing to ask whether the same holds true for compositional linguistic stimuli — sentences. In a recognition memory experiment with responses from 443 participants to 2500 six-word-long target sentences, we found that a sentence's semantic distinctiveness – as estimated through contextual representations from a large language model – predicts the accuracy and speed of its recognition. These effects were observed for both intrinsic sentence memorability (distinctiveness of a sentence relative to a large corpus of sentences) and contextual memorability (distinctiveness relative to recently encountered sentences in the experiment), and cannot be reduced to properties of the sentence's constituent words. Our findings suggest that sentence memorability, similar to image and word memorability, is related to meaning distinctiveness, thus extending the noisy representation hypothesis to compositional linguistic stimuli.