Hierarchical attention and transformers for automatic movie rating

The MPAA rating provides a guide for parents to decide if a movie is suitable for their children, and determines who is allowed into movie screenings. If the assigned rating does not match with that intended by the movie makers, the movie has to go through extra changes. Predicting this rating from the movie scripts would allow for the changes to be done even before the shooting starts, when they are the cheapest. Furthermore, automatizing this reviewing process would allow for cheaper large scale classification of videos from other sources, such as social media and streaming platforms. In this paper we propose RNN and Transformer based hierarchical architecture well suited to analyze movie scripts as large text sequences. The proposed RNN architecture outperforms the State-of-the-art (SOTA) by around 3 points in the F1 score, while our Hierarchical Transformer outperformed the SOTA in around 5 points. Furthermore, we devise a visualization strategy to address the problem of interpretability of transformers, which is particularly hard for large sequences.