Exploring Natural Language Models' Responses to Intersectional Identities in Climate Change Education (Poster 24)

• In Event: AERA Poster Session Seven
  In Poster Session: Learning in Digital Environments

Abstract

Employing natural language models (models trained on large text corpora to produce human-like text) in open-ended exchange is challenging, because the models may replicate biases from training data. In this work, we explore how three models (DialoGPT, GPT-3, BlenderBot) account for intersectional identities in discussions about climate change. Attending to the intersection of age, gender, ability, socioeconomic, and race/ethnicity is crucial in climate education, as these factors interrelate to influence climate issues' perceptions. We found that most models’ responses were overly sympathetic or uncertain, spread stereotypes, or ignored the speakers’ identities. We also found models’ responses that incorporated intersectional identities in meaningful ways. Findings highlight the need for human validation and participatory design to iteratively improve models’ data and output.

Author

• Ha Nguyen, Utah State University
  Presenting Author