Siamese text classification network (SiamTCN) for multi-class multi-label information extraction of typhoon disasters from social media data

Le résumé fourni par la source

Accurately monitoring disaster effects is a crucial task in relief efforts (e.g. typhoon rescue). Social media data plays a vital role in disaster management, while deep learning-based methods gain more attention in typhoon disaster research. However, most existing classification methods for typhoon disasters are limited to multi-class but single-label levels, contradicting the reality that a social media text may correspond to multiple types of disaster damage. This paper proposes a siamese text classification network (SiamTCN) for multi-class multi-label information extraction from typhoon disasters based on Sina Weibo data. The SiamTCN leverages a dual-path architecture with shared weights, utilizing multi-head self-attention and convolution to extract hidden features from texts. A novel multi-class multi-label contrastive loss function is proposed to optimize the model. Additionally, address information is extracted through address matching and check-in locations. The spatio-temporal characteristics provide actionable insights for disaster management, enabling timely and targeted responses to affected regions. Experiments are conducted on Sina Weibo texts collected from six typical typhoon land-falls in Chinese coastal regions from 2018 to 2023. The results demonstrate that the accuracy achieved by the proposed method is 0.9454, 0.9391, and 0.9422, respectively. The code for this paper is available at https://github.com/SiamTCN.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.