Social Networks for Natural Disaster: Operational Interpretation – ReSoCIO

Because of the extent of the scientific challenges addressed, the project focuses on two types of natural hazards studied using data from the social network Twitter alone, chosen for its ease of access to data that are widely disseminated: 1. flash floods, which represent a major risk that is difficult to anticipate and whose frequency is increasing, particularly in southern France, and 2. earthquakes, which constitute a major unpredictable and potentially very destructive risk that threatens a significant part of the national territory. Covered by the recognized thematic expertise of BRGM and PREDICT-Services, these hazards are the result of very different processes giving rise to crisis situations with varied geographical coverage and characteristics. Regarding the tools developed within the framework of the project, it is proposed to set up interoperable webservices, efficient and flexible means to allow communication between platforms, and which do not require heavy development of user interfaces. Finally, in order to facilitate the analysis of the organizational dynamics of the network of crisis management actors, the project will concentrate its analyses on the pilot territory of South-East France, focusing on several use cases: the analysts of PREDICT-Services, Fire and Rescue Services, communal crisis cells, ... On the basis of these use cases, the project gathers a multidisciplinary team in order to meet the targeted objectives. The work program and the associated research activities are organized in six work packages (WP): - WP0: Coordination and dissemination, - WP1: Elaboration of reference events and datasets - WP2: Collection, analysis and extraction of information - WP3: Contextualization - WP4: Decision support - WP5 : Organizational analysis of uses In a research landscape with plethoric and often non-perennial platforms, the great specificity of the ReSoCIO project is to be based on 3 existing, robust and complementary platforms, allowing a rise in abstraction from the scale of raw data (SURICATE-Nat), to that of information (RIOSUITE) and up to that of useful information for decision making (WikiPREDICT).

WP1. Two applications were developed, one for the retrieval, storage and dissemination of collections of tweets between partners, and the other to allow manual labeling of tweets. Among the flash floods and earthquakes that have occurred in France in recent years, 22 events were deemed of interest to the project, to which 3 international earthquakes were added. The most notable events were the subject of a summary report describing the development of the phenomenon, its effects and the crisis management actions taken. WP2. Integration of a first version of a model for extraction and disambiguation of place names in the SURICATE-Nat platform. This is a deep learning model based on a dataset specifically built to support the extraction of geographical information from French tweets. A multilingual learning model to identify tweets related to a crisis has also been developed and will be implemented in SURICATE-Nat. Finally, a work of annotation of tweets is in progress to train new predictive algorithms. WP3. Development of a flood detection algorithm based on variations in the frequency of tweets. The 1st version of a semi-supervised information extraction algorithm has been produced, in order to identify the words of interest in the tweets. Simple logical rules allow to instantiate semantic links between words extracted from the same tweets in order to generate contextualized situational models. These rules will soon be enriched with spatio-temporal dimensions. WP4. A work of identification of the links to be established between the partners' platforms has been carried out: it will guide the creation of webservices. In addition, during the 2022 steering committee, representatives of emergency services were able to present to the partners their feedback on the use of social media in support of the emergency situation. A workshop was also organized to identify the needs of a number of crisis management practitioners in this area. WP5. An exploratory analysis of the inter-organizational context was conducted with interviews that highlighted the information needs of practitioners as well as to better understand their understanding of social networks. A case analysis was initiated to explore the flash floods of the Roya and Vésubie valleys in October 2020. Qualitative data was collected via LT1 (RETEX, commission report) as well as through interviews with experts.

Webservices delivering real-time results from the analysis of tweets are currently being developed. These webservices will be integrated into the PREDICT-Services crisis hypervisor in 2023, so that they can be tested in an operational context. Thanks to this mechanism, the ReSoCIO project will enter a phase of increased involvement of crisis management practitioners, with in particular the implementation of an experimental design from spring 2023, allowing to evaluate the contribution of the tool within organizations.

Adrot A. et al. (2022, May). Using Social Media Data in Emergency Management: a proposal for a socio-technical framework and a systematic literature review. In ISCRAM 22.
Adrot, A., Aguerre, M., Data Ecosystems and Disaster Risk Reduction in Cross-border Regions: Visioning from 2020 Roya Valley Flood Disaster. ISCRAM, 2022, Tarbes (France)
Auclair, S. (2020, November). Prospects for better disaster management using AI. Human-centric AI – 2nd French-German-Japanese Symposium, “AI & Environment: Risk prevention” Session
Caillaut, G., Gracianne, C., Abadie, N., Touya, G., & Auclair, S. (2022, May). Automated construction of a French Entity Linking dataset to geolocate social network posts in the context of natural disasters. In ISCRAM 22. hal.archives-ouvertes.fr/hal-03631387/
Caillaut, G., Gracianne, C., Auclair, S., Abadie, N., & Touya, G. (2022, June). Annotation sémantique pour la géolocalisation d'entités spatiales dans des tweets. In PFIA Résilience et IA. hal.archives-ouvertes.fr/hal-03682484/

Two-thirds of French municipalities are now considered to be exposed to natural disasters. The risk of flooding is the most frequent, while seismic risk is among the most feared in terms of the number of potential victims. This situation is likely to increase in the coming decades, due to the combination of an increase of both the frequency and intensity of certain hazards due to climate change, with demographic dynamics that tend to concentrate populations in areas at risk, especially along the coastline. As a result, and with significant territorial heterogeneity, the authorities have to deal with the diversity of hazards to establish efficient crisis management systems able to protect people and property. This requires, first and foremost, the construction of situational awareness to better target and anticipate decision-making, and requires the multiplication of feedback channels from the field. In this context, RéSoCIO aims to propose a tooled approach to demonstrate the interest and feasibility of the automated exploitation of data from the Twitter social network in the context of crises related to fast-kinetic natural disasters, focusing on the case of flash floods and earthquakes. This objective requires overpassing many specific locks, the first of which concerns the ability, in a context of natural disaster, to provide information quickly. The lifting of this lock is all the more delicate as the major societal challenge of a more targeted and rapid response to natural disasters requires tweets to be analyzed with other types of exogenous data that are rapidly available (releases of weather forecast models, earthquake characteristics, etc.), and that the analysis process is automated to allow continuous "on-the-fly" processing. Another type of lock to be lifted is related to the acceptability of this type of contribution and its potential exploitability in real context. It stems from the uncertain nature of the adoption of innovative tools - due to difficulties in use and the cultural context that can lead to their abandonment - which can only be reduced by taking into account the organizational context. On this basis, RéSoCIO offers a cross-sectional analysis of how data from Twitter can be exploited by actors involved in the operational management of natural disasters, with the main innovative feature of coupled identification of algorithmic and organizational mechanisms. The proposed work is divided into different stages: monitoring, analysis, and then merging of tweets with other data and knowledge. In a research landscape with plethoric and often non-sustainable platforms, the great specificity as well as the differentiating point of the project is to rely on 3 existing, robust and complementary platforms, allowing a rise in abstraction from the scale of raw data ( SURICATE-Nat), that of information (RIOSUITE) and that of information useful for decision-making (WikiPREDICT). The development of these 3 platforms alone proves demonstratively the consideration, and the contribution to the state of the art: they will become the natural receptacles of the methodologies of analysis developed in the project.