Topic detection on social media and online news

• Main Author: Li, Yuhao
• Other Authors: Mao Kezhi
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/152938
• Institution: Nanyang Technological University

In recent years, more and more people like to publish and obtain information on the Internet through various mobile phone applications, and this information mainly exists in the form of short texts. For example, when some natural disasters occur, someone has a Twitter or Facebook message. This message may be about or at the same time about "flood", "storm" and "earthquake". Use intelligent technology to classify these short texts for timely detection The occurrence of disasters and the judgment of disaster types are of great significance. Because of its automatic learning and high-dimensional features, deep learning technology has attracted many scholars to apply it to the field of text analysis. Therefore, this paper has carried out research on the text classification technology based on deep learning, and the main work is summarized as follows. In the first place, the research status of text classification technology at home and abroad is investigated, and the main technologies, which are based on deep learning, involved in text classification process are introduced. From text preprocessing and presentation, to CNN, RNN, Attention mechanism, and at last to models based on transformer. In view of the small number of disaster-related public social network message text datasets, a web crawler is used for data crawling, a new dataset is built, which is labeled for multi-class and multi-label classification task as well as the existing public datasets. Then, a deep learning-based text classification model CRNNA is established. The GloVe model is used to preprocess the text, and then the high-dimensional features were extracted through the convolutional neural network, and then the bidirectional LSTM is used in order to capture the contextual features and global features. In addition, Hierarchical Attention mechanism is introduced combined with LSTM layer, and finally the output is classified. Experiments have proved that it has achieved better performance than traditional methods on both multi-class and multi-label classification tasks.