Classifying Natural Disaster Tweet using a Convolutional Neural Network and BERT Embedding

Social media platforms have become a medium to find a vast source of information throughout the internet. Twitter has become one of the more popular microblogging platforms out there, and the more users there are in these platforms means the more various types of information can be sent out in a day. On Twitter users are able to write their expression in the form of tweets, this will then create a post on twitter's timeline and other users are able to see these tweets. If a tweet suddenly gets viral, Twitter will put the user's tweets into the trending page allowing even more users to view the said tweet. During an event of a natural disaster often a lot of the tweets that are being posted, have mention of the disaster making it a trending topic on Twitter. From this, a vast number of tweets about a disaster can be collected as data, but not always are the tweets containing information about the disaster. Often there are tweets that use natural disaster words but do not actually talk about the disaster itself, hence are not informative and can be classified as a non-disaster tweet. This research paper aims to propose a system to classify the disaster tweets and the non-disaster tweet during a disaster. The proposed method is based on Convolutional Neural Network (CNN), using a Bidirectional Encoder Representation from Transformers (BERT) as an Embedding. As a comparison, it will then be compared with another embedding method named Word2Vec. The Evaluation result after training and testing of the CNN with BERT embeddings gave the most consistent results attaining accuracy of 97.16% precision of 97.63%, a recall of 96.64, and an f1-score of 97.13% for the model classification.