Reading between the lines: A hybrid RNN architecture to detect prejudice in movies

Abstract

Watching movies is considered the most popular pastime today. Keeping in mind research that has indicated that adolescents imbibe social behavior from movies, we propose a hybrid deep learning model to detect the degree of prejudice (sexism, racism, ableism, xenophobia, homophobia etc) in movies. Using a stacked bi-LSTM and XGBoost model to make predictions, we compared the percentage of prejudice in movies from two time periods (1975-2000 and 2001-2020). Further, we compared animated movies from production houses like Disney, Pixar with movies often rated R or PG-18 by Motion Picture Content Rating Boards. The results showed that the percentage of prejudice has been significantly reduced in the newer decades, and that, surprisingly, movies intended for a younger target audience, on average, perpetuate more prejudice than those meant for late teenagers and adults.