2021 Honors Thesis: Junaid Ahmed

Name:

Junaid Ahmed

Title of Thesis:

Can NLP Models Aid in Behavioral Economics Decision-Making?

Committee:

Stephen O'Connell, Mike Carr (Mathematics), Phillip Wolff (Psychology)

Abstract:

Natural Language Processing (NLP) models have seen rapid improvements in the last two years. Literature has indicated that these models are capable of reasoning, and in certain cases, reason better than humans. While behavioral economics tends to focus exclusively on human subjects, this study seeks to evaluate how NLP models fare in comparison to human subjects during cognitive bias tasks. More specifically, we evaluate how RoBERTa responds to fill-in-the-blank questions based on the conjunction fallacy. We use the conjunction fallacy due to its mathematical falsifiability and ease-of-testing. The hypothesis guiding this study is that RoBERTa outperforms human subjects. From this study, we conclude that RoBERTA does not outperform human subjects in aggregate, but shows promise for individuals prone to the conjunction fallacy, suggesting that there is value in future research. Moving forward, we recommend that other NLP models undergo similar tests across a greater range of cognitive biases to more accurately assess whether there is potential for using NLP models as external aids to decision making.