Modeling Sentiment–Hydrology Interaction Using LLM: Insights for Adaptive Governance in Ceará’s Water Management

Abstract

This study aims to analyze the relationships between concerns and sentiments of stakeholders and the drought stage in a semi-arid region of Ceará from Language Technologies based on Artificial Intelligence. The dataset comprises 36 meeting minutes of water management bodies (2007–2024), of which 17 correspond to dry periods and 19 to normal periods (reservoir volume > 50%). Natural Language Processing (NLP) techniques were applied to generate word clouds, and sentiment analysis was performed using a Large Language Model (Llama 3.2, 3B). Sentiment scores were compared with reservoir volume data. Results show that both perceptions and themes differed between drought and normal phases, with higher water availability coinciding with more positive sentiments. A moderate positive correlation was found between sentiment and reservoir volume (r = 0.53, p = 0.00095, 95% CI [0.24, 0.73]). Statistical tests confirmed differences between periods (Welch’s t-test, p = 0.0018; Mann-Whitney, p = 0.0039). Box-plot analyses indicated that over 75% of sentiments were positive in normal phases, while about 65% were negative in drought phases. These findings highlight the sensitivity of human perceptions to hydrological conditions and point to the potential of LLMs as innovative instruments for integrating qualitative data into complex socio-environmental analyses.