Abstract:Social media users expect quick and high-quality responses from emergency services when seeking help. However, these organizations face difficulties in detecting and prioritizing critical requests due to the overwhelming amount of information on social media and their limited human resources to tackle it during mass emergencies or disaster events. The situation is exacerbated when users communicate in different or native languages, which can be expected during disasters. While recent studies have focused on characterizing and automatically detecting help requests on social media, they focused on non-behavioral features and monolingual data, primarily in English. Thus, a key gap exists in analyzing multilingual requests on social media for public services. In this paper, we introduce a knowledge distillation framework called MulTMR (Multiple Teachers Model for detecting and Ranking), which combines the power of both task-related and behavior-guided models as diverse teachers for training a student model to efficiently detect serviceable request messages across languages and regions on social media during natural disaster events. We demonstrate that the presented framework can enhance performance (with an AUC improvement of up to 10%) in various scenarios of multilingual test data. Our results, which were validated on real-world data collected in three languages during ten disasters across seven countries, indicate the use of behavior-guided teacher models in MulTMR can increase attention to relevant indicators of serviceability characteristics. The application of the MulTMR framework through a streaming data analytics tool could reduce the cognitive load on personnel within social media teams of emergency services. Further, its application could inform how to leverage human behavior characteristics in creating automated models for social media analytics to design systems in other public service domains beyond emergency management.

Multi-task Multimodal Learning for Disaster Situation Assessment.

Multimodal Deep Learning Based on Multiple Correspondence Analysis for Disaster Management

Multi-modal deep learning framework for damage detection in social media posts

Disaster Image Classification by Fusing Multimodal Social Media Data

Disaster assessment from social media using multimodal deep learning

Dynamic Task and Weight Prioritization Curriculum Learning for Multimodal Imagery

Multimodal Social Sensing for the Spatio-Temporal Evolution and Assessment of Nature Disasters

A Multimodal Data Analysis Approach to Social Media during Natural Disasters

Multimodal tweet classification in disaster response systems using transformer-based bidirectional attention model

Enhancing multimodal disaster tweet classification using state-of-the-art deep learning networks

A Social Context-aware Graph-based Multimodal Attentive Learning Framework for Disaster Content Classification during Emergencies

A Multimodal Resource Balancing Approach Based on Cross-Attention and Bidrectional Associative Memory Networks

Transformer-based Multi-task Learning for Disaster Tweet Categorisation

MS4D-Net: Multitask-Based Semi-Supervised Semantic Segmentation Framework with Perturbed Dual Mean Teachers for Building Damage Assessment from High-Resolution Remote Sensing Imagery

Harnessing Diverse Data for Global Disaster Prediction: A Multimodal Framework

Real-Time Classification of Disaster Images from Social Media with a Self-Supervised Learning Framework

Multimodal Deep Representation Learning for Video Classification

CrisisSense-LLM: Instruction Fine-Tuned Large Language Model for Multi-label Social Media Text Classification in Disaster Informatics

Robust Training of Social Media Image Classification Models for Rapid Disaster Response

Collaborative Emergency Decision-Making: A Framework for Deep Learning with Social Media Data

Multilingual Serviceability Model for Detecting and Ranking Help Requests on Social Media during Disasters