With the vast amount of social media posts available online, topic modeling and sentiment analysis have become central methods to better understand and analyze online behavior and opinion. However, semantic and sentiment analysis have rarely been combined for joint topic-sentiment modeling which yields semantic topics associated with sentiments. Recent breakthroughs in natural language processing have also not been leveraged for joint topic-sentiment modeling so far. Inspired by these advancements, this paper presents a novel framework for joint topic-sentiment modeling of short texts based on pre-trained language models and a clustering approach. The method leverages techniques from dimensionality reduction and clustering for which multiple algorithms were considered. All configurations were experimentally compared against existing joint topic-sentiment models and an independent sequential baseline. Our framework produced clusters with semantic topic quality scores of up to $0.23$ while the best score among the previous approaches was $0.12$. The sentiment classification accuracy increased from $0.35$ to $0.72$ and the uniformity of sentiments within the clusters reached up to $0.9$ in contrast to the baseline of $0.56$. The presented approach can benefit various research areas such as disaster management where sentiments associated with topics can provide practical useful information.

Natural disasters are increasingly prevalent worldwide, necessitating the utilisation of diverse datasets for effective disaster response. While geo-social media data represents a valuable resource in this context, the recent restrictions to Twitter data have significantly impacted its availability for disaster research. Alternative social media platforms to Twitter remain underexplored, leading to limited understanding of their potential. To address this gap, we collected posts for a specific use case, Hurricane Ian, from four social media platforms (Mastodon, Reddit, Telegram, and TikTok), and subsequently geoparsed each post. We then computed spatial and temporal patterns and evaluated their correlations to investigate the potential applicability of other data sources for disaster response efforts. While none of the platforms can fully substitute Twitter’s role in disaster management, the findings demonstrated that substantial amounts of potentially valuable data can be sourced from other platforms. Despite consistent overall patterns, subtle differences in temporal activity and spatial distribution suggest that each platform offers unique insights that enhance situational awareness. However, a significant challenge in using these platforms for disaster response is the low spatial accuracy achievable through geoparsing.

The recent COVID-19 pandemic has underscored the need for effective public health interventions during infectious disease outbreaks. Understanding the spatiotemporal dynamics of urban human behaviour is essential for such responses. Crowd-sourced geo-data can be a valuable data source for this understanding. However, previous research often struggles with the complexity and heterogeneity of such data, facing challenges in the utilisation of multiple modalities and explainability. To address these challenges, we present a novel approach to identify and rank multimodal time series features derived from mobile phone and geo-social media data based on their association with COVID-19 infection rates in the municipality of Rio de Janeiro. Our analysis spans from April 6, 2020, to August 31, 2021, and integrates 59 time series features. We introduce a feature selection algorithm based on Chatterjee’s Xi measure of dependence to identify relevant features on an Área Programática da Saúde (health area) and city-wide level. We then compare the predictive power of the selected features against those identified by traditional feature selection methods. Additionally, we contextualise this information by correlating dependence scores and model error with 15 socio-demographic variables such as ethnic distribution and social development. Our results show that social media activity related to COVID-19, tourism and leisure activities was associated most strongly with infection rates, indicated by high dependence scores up to 0.88. Mobility data consistently yielded low to intermediate dependence scores, with the maximum being 0.47. Our feature selection approach resulted in better or equivalent model performance when compared to traditional feature selection methods. At the health-area level, local feature selection generally yielded better model performance compared to city-wide feature selection. Finally, we observed that socio-demographic factors such as the proportion of the Indigenous population or social development correlated with the dependence scores of both mobility data and health- or leisure-related semantic topics on social media. Our findings demonstrate the value of integrating localised multimodal features in city-level epidemiological analysis and offer a method for effectively identifying them. In the broader context of GeoAI, our approach provides a framework for identifying and ranking relevant spatiotemporal features, allowing for concrete insights prior to model building, and enabling more transparency when making predictions.