Predicting Disaster Tweets using Enhanced BERT Model

Twitter has become an essential means of communication in emergency situations. The widespread use of smart phones, anyone can immediately report any incident they witness. As a result, more agencies are interested in programmatically monitoring Twitter. Social media sites such as Twitter provide a common forum for sharing and exchanging personal experiences with others. Most of the people share personal experiences, neighborhood problems also the activities on social websites to let unwanted peoples to know. Many rescue organizations constantly monitor this kind of data to detect disasters and reduce the likelihood of casualties. However, humans cannot manually comb through vast amounts of data to uncover threats in real-time. To achieve this, many studies have proposed representing words in a form that computers can understand and applying machine learning techniques to the word representations to determine the sentiment of the text. Previous search techniques provide single embeddings or representations of words from a given document. However, the new Advanced Contextual Embedding is Bidirectional Encoder Representations from Transformers (BERT) method creates different vectors for the same expression with different settings. Although there are no specific studies on how BERT embeddings can help analyze catastrophe-type tweets, these representations have been successfully used in various Natural Language Processing (NLP) tasks.