A Sustainable Way Forward: Systematic Review of Transformer Technology in Social-Media-Based Disaster Analytics

[18] Flood (DKI Jakarta, Indonesia) BERT-MLP BERT High accuracy (82%) in classifying tweets related to flood events using geospatial data. Stemming process may remove important features, affecting accuracy. [20] Wildfires in the Western United States (2020) BERTopic, NER BERT Real-time estimation of wildfire situations through social media analysis for decision support. Potential noise from broad search terms, inaccuracies in user location data, and single-topic document assumption. [21] Twitter-Based Disaster Prediction Improved BERT model, LSTM, GRU BERT Demonstrates superior accuracy in disaster prediction on Twitter by analyzing patterns associated with various types of disasters. Outperforms traditional models like LSTM and GRU in predicting disaster-related tweets. Not specified. [22] Epidemics, Social Unrest, and Disasters Enhanced BERT model, GloVe for feature extraction, LSTM for classification BERT Superior in terms of accuracy, precision, recall, and F1-score, capturing tweet semantics to accurately identify trend-related tweets. Not specified. [23] Detection and Classification of Natural Disasters from Social Media FNN, CNN, BLSTM, BERT, DistilBERT, Albert, RoBERTa, DeBERTa BERT and its variants Demonstrated the effectiveness of deep learning methods in accurately detecting and classifying disaster-related information from tweets, with preprocessing and bias mitigation enhancing performance. The complexity of models and the need for extensive preprocessing and bias mitigation to handle the diverse and noisy nature of social media data. [24] Various Natural Disasters (Wildfires, Hurricanes, Earthquakes, Floods) RoBERTa for text analysis, Vision Transformer for image understanding, Bi-LSTM for sequence processing, and attention mechanism for context awareness RoBERTa, Vision Transformer Superior performance with accuracy levels ranging from 94% to 98%, effective combination of textual and visual inputs through multimodal fusion. Requires substantial computational and memory resources, potential hardware limitations. [25] Various Natural Disasters (Earthquakes, Floods, Hurricanes, Fires) DenseNet, BERT BERT for textual features, DenseNet for image features Achieved an accuracy of 85.33% in classifying social media data into useful and non-useful categories for disaster response, outperforming state-of-the-art techniques. Not specified, but potential issues could include the complexity of integrating multimodal data and the need for substantial computational resources. [26] COVID-19 Pandemic Sentiment Analysis RETN model, BERT-GRU, BERT-biLSTM, nature-inspired optimization techniques GPT-2, GPT-3 Enhanced performance in sentiment analysis on large-scale datasets, including text, images, and audio. Complexity in implementation and optimization. [27] Cyclone-Related Tweets BERT, machine learning, and deep learning classifiers BERT BERT model achieves better results than other ML and DL models even on small labelled datasets. Tweets often contain ambiguity and informal language that are hard for the machine to understand [28] Various (e.g., Earthquakes, Typhoons) RACLC for classification, RACLTS for summarization BERTweet (a variant of BERT for Twitter data) High performance in disaster tweet classification and summarization. Provides interpretability through rationale extraction, enhancing trust in model decisions. Potential limitations include the need for extensive training data and the challenge of adapting to new or unforeseen disaster types. [29] Various (e.g., Shootings, Hurricanes, Floods) Transformer-based multi-task learning (MTL) BERT, DistilBERT, ALBERT, ELECTRA Demonstrates superior performance in classifying and prioritizing disaster-related tweets using a multi-task learning approach. Allows for effective handling of large volumes of data during crises. Challenges include handling the high variability of disaster-related data and the computational demands of processing large datasets in real-time. [30] Disaster Detection on Twitter BERT (Bidirectional Encoder Representations from Transformers) with keyword position information BERT Improved accuracy in disaster prediction on Twitter by incorporating keyword position information into the BERT model. Relies heavily on the keyword position, which may not always accurately reflect the context or importance of a tweet. [31] General Disaster Management Various BERT-based models (default BERT, BERT + NL, BERT + LSTM, BERT + CNN) BERT Effective in classifying disaster-related tweets by using balanced datasets and preprocessing techniques. Challenges with imbalanced data and processing informal social media text. [32] Various Crises and Emergencies Detected via Social Media MobileBERT for feature extraction, SSA improved with MRFO for feature selection MobileBERT High accuracy and efficiency in detecting and classifying crisis-related events on social media, leveraging advanced transformer technology and optimized feature selection. Challenges include computational demands for processing and analyzing large-scale social media data in real-time and adapting to diverse and evolving crisis scenarios. [33] Flood-Related Volunteered Geographic Information (VGI) BERT with TF-IDF, TextRank, MMR, LDA BERT Provides an ensemble approach combining BERT with traditional NLP methods for enhanced topic classification accuracy in flood-related microblogs. Complexity in integrating multiple algorithms and potential challenges in scalability and real-time processing. [34] Electricity Infrastructure BERT BERT for text classification Capable of sensing the temporal evolutions and geographic differences of electricity infrastructure conditions through social media analysis. Limited capability in areas with few Twitter activities, reliance on geotagged tweets, which are a small portion of total tweets. [35] Transportation Disaster Detection and Classification in Nigeria BERT with AdamW optimizer BERT Improved accuracy in identifying and classifying transportation disaster tweets with an accuracy of 82%, outperforming existing algorithms. Relies on named entity recognition (NER) for location identification, which may not be effective if users do not specify their location accurately. [36] Disaster Prediction from Tweets GloVe embeddings for word representation, BERT for classification BERT Achieved 87% accuracy in classifying tweets related to disasters, showing BERT’s superiority over traditional models like LSTM, random forest, decision trees, naive Bayes. Requires significant preprocessing and understanding of NLP concepts to implement effectively. [37] Natural Disaster Tweet Classification CNN with BERT embedding BERT Achieved high accuracy (97.16%), precision (97.63%), recall (96.64%), and F1-score (97.13%) in classifying natural disaster tweets. Requires complex preprocessing and might overfit after certain epochs, indicating a need for careful model training and validation setup. [38] Analysis and Classification of Disaster Tweets from a Metaphorical Perspective BERT, RoBERTa, DistilBERT BERT, RoBERTa, DistilBERT Demonstrated improved performance in classifying disaster-related tweets, including those with metaphorical contexts, highlighting the models’ ability to capture metaphorical text representations effectively. The study did not specifically address the computational efficiency or potential limitations in processing metaphorical content across diverse disaster types and languages. [39] Various (e.g., Earthquakes, Floods, Shootings) Transformer-based model with multitask learning approach, including a fine-tuned encoder based on RoBERTa and transformer layers as a task adapter RoBERTa Significant improvements in classifying and prioritizing tweets in emergency situations by leveraging entities, event descriptions, and hashtags. This approach benefits from the adaptability of Transformers to handle noisy social media data. The complexity of the model requires substantial computational resources for training and fine-tuning. The effectiveness of the model is dependent on the quality and representation of the input data, including the preprocessing of hashtags and the augmentation with event metadata. [40] Various (e.g., Earthquakes, Floods) BERT, GRU, LSTM BERT BERT achieved the highest accuracy (96.2%) in classifying disaster-related tweets, indicating its effectiveness in understanding and categorizing disaster information from social media. Increased complexity of the BERT architecture may lead to overfitting and requires careful adjustment.