Scholarlink

<p>Classification of disaster events is an essential task to send early warnings and informing the aid agencies and humanitarian organizations to take immediate action. Graph neural network (GNN) has recently drawn more interest for its use in the text classification task. The current graph-based models did not consider the syntactical structure and word order information. The GNN-based models build a corpus-level graph that includes a test set and needs reconstruction of the entire graph when new tweets are added to the corpus. In this paper, we propose a novel hybrid model by combining the graph attention network (GAT) and the bidirectional encoder model for transformers (BERT) to overcome these issues. The results of the conducted experiments demonstrated that the proposed model performs better than all baseline models based on GNN on seven publicly available datasets in terms of weighted average F1 score and accuracy with less memory consumption and faster convergence.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Disaster related tweet classification method based on BERT and GAT

  • 期刊论文

  • 发表日期：

    2025年2月3日

查看全文
  • Nayan Ranjan Paul,
  • Rakesh Chandra Balabantaray

摘要

Classification of disaster events is an essential task to send early warnings and informing the aid agencies and humanitarian organizations to take immediate action. Graph neural network (GNN) has recently drawn more interest for its use in the text classification task. The current graph-based models did not consider the syntactical structure and word order information. The GNN-based models build a corpus-level graph that includes a test set and needs reconstruction of the entire graph when new tweets are added to the corpus. In this paper, we propose a novel hybrid model by combining the graph attention network (GAT) and the bidirectional encoder model for transformers (BERT) to overcome these issues. The results of the conducted experiments demonstrated that the proposed model performs better than all baseline models based on GNN on seven publicly available datasets in terms of weighted average F1 score and accuracy with less memory consumption and faster convergence.