Written language is a multimodal system that integrates visual, phonological, and semantic information. This study examines whether orthographic visual distinctiveness—the degree to which word forms differ visually—acts as a structural constraint across languages. Using standardized script renderings from 131 languages, we extracted visual features of words through a Vision Transformer (VIT) and compared visual distances between co-occurring word pairs from natural corpora and random word pairs from lexicons, controlling for word length and related factors. The results show that co-occurring words are visually more distinct than expected by chance, and this effect is consistent across diverse writing systems. These findings indicate that visual distinctiveness contributes independently to the organization of written language, reflecting an underlying pressure toward visual discriminability in lexical form. Beyond linguistic implications, the framework demonstrates how deep vision models can capture cognitively meaningful visual features of text, offering new perspectives for multimodal research on orthography, reading, and cross-lingual modeling.
Wang et al. (Thu,) studied this question.