Abstract

Here we explore whether ChatGPT can provide definitions for novel derived words, and to what extent it is able to capture their degree of perceived meaningfulness, as estimated by human speakers. Whereas the produced definitions are largely sensible, ChatGPT predictions about meaningfulness do not align well with the human ones: not only ChatGPT is outperformed by an older, much simpler, cognitively-oriented model; its predictions are also shown to mostly rely on affixes, not considering the actual combinations of morpheme meanings, making it at odds with human speakers. Novel words represent an interesting challenge for large-language-models.