Cross-Language Typicality Effects in a Multilingual Large Language Model

The typicality effect is the finding that some members of a category are more ``central'' and others more ``peripheral''. This effect is seminal for understanding the mental representation of concepts. Recently, researchers have looked for typicality effects in the representations learned by machine learning models as evidence of their cognitive alignment. Studies of the typicality effect in Large Language Models (LLMs) have focused on models trained on English corpora and category norms collected from English speakers. Here, we use existing norms to investigate the typicality effect across five languages: English, French, Portuguese, German, and Spanish. We focused on eight categories common across these norms, and asked whether a multilingual LLM, GPT-4o-mini, shows human-like typicality effects across these languages. The results show variation in typicality gradients across languages. Importantly, GPT-4o-mini's typicality judgments show strong alignment with human norms for some languages: English and French. The strong performance for French, in particular, cannot simply be attributed to the representation of that language in the training corpus. We discuss the implications of these findings for future studies exploring alternative model prompting approaches, different languages, and the modeling of new category norms collected using uniform methods.