description.md

Gender Bias in MT: Hungarian to English

The Hungarian language has no grammatical gender and words like “he” and “she” are both translated as “ő”. This makes it an interesting language to study gender bias in machine translation (MT) models, when translating to another language that does distinguish between “he” and “she”. In this demo, we will test the OPUS-MT models (Tiedemann & Thottingal, 2020) from the Language Technology Research Group at the University of Helsinki (Helsinki-NLP).

For each translation, we also use the Inseq library to compute the feature attributions with integrated gradients: How important is each token in the source (Hungarian) for the translation of the target tokens (English)?

⚠️ Please note that this demo is just an illustration of how gender bias could manifest in MT models, but an actual assessment of its bias requires a more rigourous experiment.