This research project investigates the effectiveness of sentiment analysis tools on tweets discussing neurodivergent individuals, particularly those with autism. Traditional models like TextBlob often lack the contextual awareness needed to interpret subtle or emotionally complex content. To address this, we developed a system comparing TextBlob and GPT-4 using both classification and regression-based evaluation 1. A dataset of 100 tweets was analyzed. In the first experiment, GPT-4 achieved a macro F1-score of 0.61, outperforming TextBlob’s 0.58, with both models reaching 62% accuracy. In the second experiment, which evaluated polarity scoring, GPT-4 achieved a MAE of 0.604, RMSE of 0.766, and a correlation of 0.479, compared to TextBlob’s MAE of 0.650, RMSE of 0.778, and correlation of 0.394. These results confirm that GPT-4 provides more accurate and context-sensitive sentiment predictions 2. This system improves upon prior lexicon-only approaches by combining classification and polarity scoring to offer a comprehensive, real-world analysis of sentiment in neurodivergent conversations.
Cui et al. (Sat,) studied this question.