Abstract
This thesis investigates gender and sentiment in Norwegian book reviews and their impact on machine learning models trained on the book review data. Our analysis reveals that female critics and authors give and receive significantly lower ratings than males. Using methods from interpretable machine learning, we go on to show that these statistical differences make models trained on the data associate features related to female gender with a lower sentiment than features related to male gender. We also explore the effects of gender normalization on the models' predictions and the impact of supplying models with gender knowledge during training. Our findings demonstrate the potential of interpretation methods for transformer models on Norwegian text and highlight the strengths and weaknesses of different methods for interpreting machine learning models.