Identifying Gender Bias in Generative Models for Mental Health Synthetic Data

Daniel Cabrera Lozoya, Simon D'Alfonso, Mike Conway

Abstract: Natural language generation (NLG) systems have proven to be effective tools to create domain-specific synthetic data. The mental health research field could benefit from data augmentation techniques, given the challenges associated with obtaining and utilizing protected health information. Yet, NLG systems are often trained using datasets that are biased with respect to key demographic factors such as ethnicity, religion, and gender. This can perpetuate and propagate systematic human biases that exist and ultimately lead to inequitable treatment for marginalized groups. In this research we studied and characterized biases present in the Generative Pre-trained Transformer 3 (GPT-3), which is an autoregressive language model that produces human-like text. The prompts used to generate text via GPT-3 were based on the Brief Cognitive Behavioral Therapy framework, and each prompt also specified to write the answer as a female or male patient. By controlling the sex distributions within our prompts, we observed the impact of each trait in the generated text. The synthetic data was analysed using the Linguistic Inquiry and Word Count software (LIWC-22) and ccLDA for cross-collection topic modeling. LIWC-22 results show that stereotypical competence features such as money, work, and cognition are more present in the male’s synthetic text, whereas warmth features such as home, feeling, and emotion are highly present in female’s generated data. The ccLDA results also associate competence features with males and warmth features with females.