Proceedings of the Annual Meeting of the Cognitive Science Society

This study examined whether grammatical gender, a structural property of language, causally shapes word representations to exaggerate implicit biases about social gender. Because isolating the effect of grammatical gender from semantic, socio-cultural, and (other) syntactic factors is difficult in human studies, we used word embeddings. We created 97 near-English languages from an English corpus, holding content constant but varying gender marking systems (e.g., which genders were marked, which parts of speech were marked). Separate FastText embeddings were trained on each corpus. Introducing grammatical gender increased implicit gender bias for nouns with grammatical gender but no inherent gender (grammatically-feminine nouns became more associated with women, and vice versa for grammatically-masculine nouns). This effect strengthened as agreement extended to more parts of speech and was strongest under asymmetric marking. These results provide causal evidence that grammatical gender alone may amplify gender-related social biases of word representations in artificial and, perhaps, human minds.