Can Official Data be Trusted? Clarifying Biases in Sentiment Analysis

High-quality training data are undoubtedly essential for developing machine learning models and accurately interpreting their outcomes. However, due to the psychological biases and inconsistencies inherent in human assessments, relying solely on human labelling is inadvisable. This study explored the inconsistencies and psychological biases in human economic assessments using insights from behavioral economics and empirical analysis of real-world data. The observed correlation between declining economic indicator values and increased response counts highlighted heightened engagement during periods of economic uncertainty. Moreover, advanced computational techniques revealed a divergence exceeding 60% between human-labelled and machine-classified judgments. These findings underscored the presence of significant biases that reduce the reliability of official indicators and existing sentiment models. Consequently, the study emphasized the importance of employing nuanced analytical frameworks to correct subjective distortions and enhance the robustness of sentiment analysis.