Urgency Detection in Social Media Texts Using Natural Language Processing

Abstract

Social media platforms are used extensively to request emergency help during large-scale natural or humanitarian disasters. Machine learning can be used best to mobilize resources in any disaster. Natural language processing comes to our rescue to help detect urgency, identifying certain expressions, words, and phrases that typically showcase concern or immediate need. A deep learning model can be implemented, which can be used to classify the messages as a particular urgent class label. First responders or governments can use this urgency class label to categorize the requests to provide relief measures to needy people. We have used the hugging face disaster messages dataset and reclassified messages based on urgency. We leveraged the advantages of both CNN and LSTM to get high-level features and long-term dependencies. We developed a model combining Convolutional neural networks (CNN) and bidirectional long short-term memory(BiLSTM) networks. We also conducted experiments with various model architectures, such as ensemble and hierarchical, along with varying word embeddings. We observed the performance of fine-tuning pre-trained models such as BERT and DistilBERT with our dataset. We evaluated that the developed model using CNN and BiLSTM performs better with only 10% of trainable parameters comparable to pre-trained models such as BERT, which has 110M trainable parameters. We could also observe that general-purpose word embeddings such as Glove give comparable results to domain-specific word embeddings.