Boys don't cry (or kiss or dance): A computational linguistic lens into gendered actions in film

Abstract

Contemporary media is full of images that reflect traditional gender notions and stereotypes, some of which may perpetuate harmful gender representations. In an effort to highlight the occurrence of these adverse portrayals, researchers have proposed machine-learning methods to identify stereotypes in the language patterns found in character dialogues. However, not all of the harmful stereotypes are communicated just through dialogue. As a complementary approach, we present a large-scale machine-learning framework that automatically identifies character's actions from scene descriptions found in movie scripts. For this work, we collected 1.2+ million scene descriptions from 912 movie scripts, with more than 50 thousand actions and 20 thousand movie characters. Our framework allow us to study systematic gender differences in movie portrayals at a scale. We show this through a series of statistical analyses that highlight differences in gender portrayals. Our findings provide further evidence to claims from prior media studies including: (i) male characters display higher agency than female characters; (ii) female actors are more frequently the subject of gaze, and (iii) male characters are less likely to display affection. We hope that these data resources and findings help raise awareness on portrayals of character actions that reflect harmful gender stereotypes, and demonstrate novel possibilities for computational approaches in media analysis.

Fig 1. Computational linguistic lens into gendered actions in film.

Our framework starts with a dataset of over 1.2 million movie descriptors from 912 movies and then implements three steps: first, the annotation process, where we collect manual annotations for 9,613 descriptions and over 1.5 million gender expression labels for characters. In the second step, we develop a machine learning model to identify actions, agents and patients from the natural language found in the movie scripts. Our model—trained on a large-set of general-domain documents and fine-tuned on a manually labeled description set—identifies actions, agents and patients from the natural language descriptions found in movie scripts. We use this model to automatically label the complete dataset for analysis. In the final third step, we perform a series of statistical analysis to uncover portrayal differences along characters’ portrayed attributes.

Fig 2. Labeling task.

Annotators are presented with a sentence and an action. They are asked to either select the agent (source) and patient (target) of the action. For cases where one of these is missing, the annotator has the option to check the ‘Does not say’ box.

Fig 3. Proposed SRL system.

Starting at the bottom, the input to the system is an action description in natural language. The output, shown at the top of the figure, is a sequence of labels (one per word). Labels indicate whether this word depicts an action, agent, patient or none. From its inputs, our model obtains a highly-contextualized representation for each word using the BERT transformer [53]. Each representation corresponds to a high dimensional dense vector that encodes the semantics of that word and the context it plays within the sentence. The sequence of vector representations is then fed into a recurrent neural network and a softmax layer for sequence labeling. As a post-processing step, a set of heuristics aggregate conjunctions to handle the case of groups of agents or patients.