I Simposio de Postgrado 2023. Ingeniería, ciencias e innovación

I SIMPOSIO 2023 A MULTILINGUAL AND MULTI-DOMAIN APPROACH FOR CRISIS CLASSIFICATION IN SOCIAL MEDIA ABSTRACT Social media data has emerged as a useful source of timely information about real-world crisis events. Its users can provide immediate information from the locations where events are unfolding. Based on this information, new research and tools have emerged to support crisis management. However, most of the studies in this area (known as Crisis Informatics ) have focused on a specific language (usually English) or a particular domain (type of event, such as earthquake). This limits the applicability of current approaches to new types of crises in different languages. This work aims to study how to leverage labeled data from high-resource languages and domains for addressing fundamental crisis informatics tasks in low-resource scenarios. Its main goal is to create new multilingual and multi-domain models for classifying crisis-related messages. For that, it proposes an experimental framework based on combinations of multilingual data representations (e.g.,MUSE,mBERT, XLM-R) and knowledge transfer scenarios (e.g., Monolingual & Cross- Domain, Cross-Lingual &Monodomain, Cross-Lingual &Cross- Domain, Multilingual & Multi-Domain). The experimental results show that it is possible to leverage English data to classify new crisis domains in other languages, such as Spanish and Italian (80.0% F1-score). Furthermore, this work proposes zero and few shot classifiers based on prompting generative Large Language Models. Preliminary results from this last approach are promising, especially for more specific tasks. For example, categorizing messages according to humanitarian information. Overall, this work contributes to the mitigation of cold-start situations during emergency events, when time is of essence. 1 Departamento de Ciencias de la Computación, Universidad de Chile. 2 Instituto Milenio Fundamentos de los Datos, Chile. 3 Centro Nacional de Inteligencia Artificial, Chile. 4 Amazon.com, USA. *Email: cinthia.sanchez@ug.uchile.cl Cinthia Sánchez 1,2* , Andrés Abeliuk 1,3 , Bárbara Poblete 1,2,3,4