Disaster tweet classification: A majority voting approach using machine learning algorithms

Abstract

Nowadays, people share their opinions through social media. This information may be informative or non-informative. Filtering informative information from social media plays a challenging issue. Nevertheless, people will interact more with that particular disaster event on social media, primarily when a disaster occurs. They share their opinion through some textual information such as tweets or posts. In this work, we propose a generalized approach for categorizing the informative and non-informative media on Twitter. We collected the seven natural disaster events from the crisisNLP. These datasets are different disaster events containing people’s opinions on that specific event. We pre-process the information, which converts the tweet information into machine-understandable vectors. Various machine learning algorithms have processed these vectors. We consider the individual performance of each ML algorithm on different disaster datasets upon choosing the best five algorithms for voting techniques. We tested the performance with matrices such as accuracy, precision, recall, and F1-score. We compared our results with existing models in which our proposed model performed better than other existing state of the art models.

Categorization of tweets for damages: infrastructure and human damage assessment using fine-tuned BERT model

“…The authors used lexicons and TF-IDF features with six ML models and revealed that their framework outperformed the baselines. Later, a majority voting-based approach is presented in Krishna, Srinivas & Prasad Reddy (2022) to identify only informative tweets using word2vec, TF-IDF, and the Glove model. Their model showed significant performance but they did not handle infrastructure and human damages.…”

“… Priya et al (2018) developed a query-based information retrieval method for infrastructural damage assessment but did not address human damage. Moreover, few studies proposed approaches for informational vs non-informational tweet identification like the majority voting approach ( Krishna, Srinivas & Prasad Reddy, 2022 ), and multi-model approach using image and text data ( Koshy & Elango, 2023 ) but did not address human damage assessment. According to our knowledge, only one study, Madichetty & Sridevi (2021) , focused on damage assessment for infrastructural and human damages from tweets and used a lexicon and frequency-based approach (hand-crafted) with traditional machine learning (ML) models.…”