VocabGCN-BERT: A Hybrid Model to Classify Disaster Related Tweets

Abstract

When it comes to classifying tweets about disasters, Deep Neural Network-based models have shown great potential over conventional machine learning models. In particular, Bidirectional Encoder Representations from Transformers(BERT) are effectively used to capture the contextual information present in a tweet. But it is not effectively capturing the global structural information of tweets. A Graph Convolutional Network's(GCN) power lies in its ability to capture global information. In this study, we present a novel hybrid model called VocabGCN-BERT by combining the GCN made from a vocabulary graph of tweets and the pre-trained BERT model. A powerful representation for classifying tweets is created by combining local contextual information acquired from BERT with global structural information acquired from VocabGCN. The results of the experiments demonstrate that the proposed VocabGCN-BERT performs better than the currently available state-of-art models based on GCN on seven publicly available datasets by a margin of +1.66% to +4.79% in weighted average F1 score and +1.45% to +4.34% in accuracy.