Assume My Gender: Representations of Gendered Speech in ChatGPT Outputs and Film Dialogue

Abstract

This project extends my Junior independent work by revisiting the analysis of gendered differences in movie dialogue with five times more labeled data. I employ ChatGPT-3.5 Turbo, one of OpenAI's generative AI models, to label more than 8,000 rows of new speaker dialogue by gender. In the process, I investigate the capacity of large language models to assign accurate gender labels by assessing various prompting methods. I also highlight ChatGPT's perception of gender by using Natural Language Processing (NLP) to analyze its reasoning for assigning male and female labels. Next, I use NLP techniques to investigate differences in the male and female dialogue of my new dataset. These techniques include tokenization, n-gram frequency ranking, and TF-IDF ranking. By highlighting thematic and linguistic patterns in male and female dialogue, I demonstrate the presence of gendered stereotypes in movie dialogue, thereby affirming the prevalence of these stereotypes in the real world.