Social media data analysis in a smart city context can represent an efficacious instrument to inform decision making. The manuscript strives to leverage the power of Natural Language Processing (NLP) techniques applied to Twitter messages using supervised learning to achieve real-time automated event detection in smart cities. A semantic-based taxonomy of risks is devised to discover and analyse associated events from data streams, with a view to: (i) read and process, in real-time, published texts (ii) classify each text into one representative real-world category (iii) assign a citizen satisfaction value to each event. To select the language processing models striking the best balance between accuracy and processing speed, we conducted a pre-emptive evaluation, comparing several baseline language models formerly employed by researchers for event classification. A heuristic analysis of several smart cities and community initiatives was conducted, with a view to define real-world scenarios as basis for determining correlations between two or more co-occurring event types and their associated levels of citizen satisfaction, while further considering environmental factors. Based on Multiple Regression Analysis (MRA), we established the relationships between scenario variables, obtaining a variance of 60%–90% between the dependent and independent variables. The selected combination of supervised NLP techniques leverages an accuracy of 88.5%. We found that all regression models had at least one variable below the 0.05 threshold of the f − t e s t , therefore at least one statistically significant independent variable. These findings ultimately illustrate how citizens, taking the role of active social sensors, can yield vital data that authorities can use to make educated decisions and sustainably construct smarter cities. • We use a bespoke taxonomy of risks to detect events on social media in a smart city. • As active social sensors, citizens provide valuable insight for smart city management. • We determine general patterns and trends in citizen sentiment in a smart city context. • Our Machine Learning techniques show an accuracy of 88.5% for risk event detection. • We validate relationships between user satisfaction and events by Regression Analysis.