Mirroring hierarchical attention in adversary for crisis task identification: COVID-19, Hurricane Irma

ABSTRACT

A surge of instant local information on social media serves as the first alarming tone of need, supports, damage information, etc. during crisis. Identifying such signals primarily helps in reducing and suppressing the substantial impacts of the outbreak. Existing approaches rely on pre-trained models with huge historic information as well as on domain correlation. Additionally, existing models are often task specific and need auxiliary feature information. Mitigating these limitations, we introduce Mirrored Hierarchical Contextual Attention in Adversary (MHCoA2) model that is capable to operate under varying tasks of different crisis incidents. MHCoA2 provides attention by capturing contextual correlation among words to enhance task identification without relying on auxiliary information. The use of adversarial components and an additional feature extractor in MHCoA2 enhances its capability to achieve higher performance. MHCoA2 reports an improvement of 5 - 8% in terms of standard metrics on two real-world crisis incidents over the state-of-the-art. © 2021 Information Systems for Crisis Response and Management, ISCRAM. All rights reserved.