Semantic gender associations and grammatical gender in word embeddings
George Flint, Anna Ivanova
Massachusetts Institute of Technology
To explore the relationship between grammatical gender and semantic associations, we used FastText models for English, Spanish, and German, and, in each language, surmised a list of nouns of masculine and feminine grammatical genders for languages with any. We also collected lists of adjectives adjacent in the semantic space to man or woman for semantically masculine groups and semantically feminine groups respectively in all three languages. We then collected cosine similarity data of all noun/adjective combinations within each grammatical gender group/semantic gender association group combination to test for a statistically significant effect of the grammatical gender of a noun group on its average cosine similarity to the adjective group of corresponding semantic gender association.