Twitter breaking news detector in the 2018 Brazilian presidential election using word embeddings and convolutional neural networks
Kenzo Miranda Sakiyama, André Quintiliano Bezerra Silva, Edson Takashi Matsubara · 2019
A breaking news event detector based on the time series of the number of positive, negative and neutral tweets obtained from a sentiment analysis classifier is proposed. The detector collects real-time tweets related to candidates and transforms them into word embeddings using the FastText algorithm. Using domain adaptation, the sentiment analysis classifier is trained based on a convolutional neural network (CNN) known as TextCNN. The number of positive, negative and neutral tweets in a time frame results in a time-series, which is monitored by an unsupervised time-series anomaly detector. The results show that the sentiment analysis classifier achieves an accuracy of 74% for the three classes, and the detector successfully detects significant breaking news in the 2018 Brazilian presidential election.