Multi-source domain adaptation approach to classify infrastructure damage tweets during crisis
摘要
Real-time crisis information on social media is crucial for supporting relief and rescue operations during the early stages of a crisis. However, the lack of sufficient information about ongoing incidents and the wealth of data from previous crises necessitate the use of domain adaptation (DA) techniques over other methods. Nevertheless, current DA approaches often fail to fully utilize the available past resources, resulting in the loss of important information for ongoing crises and negatively impacting performance. Existing pitfalls of state-of-the-art models are: (1) models do not work on joint domain feature relation at elementary and instance level to exploit the complete information of each domain, and (2) moreover, these models could not efficiently harness the information, when there are diversified and varying number of source crisis incidents. Inspired by the ensemble setup in identifying the infrastructure damage, we introduce Ensemble model using the elementary feature (Parts-of-speech tagging) Attention and Hypersphere Separator (EnPHyS). It operates at joint feature levels where each level works with the abundant source and scarce target data to extract the best of the (1) shared and (2) invariant features for the objective task. Ensemble uses multi-task learning (MTL) and an adversarial approach to enhance the information retrieval of target features. EnPHyS performance was investigated under single-source as well as multi-source domain adaptation scenarios with four publicly available datasets. The reported results on standard metric F-measure reveal the average growth of