Enhanced Unsupervised Data Augmentation for Emergency Events Detection and Classification

We propose to jointly detect and classify emergency events using a multi-class text classifier, which is a typical deep learning architecture with transformer modules and particularly employs Bidirectional Encoder Representations from Transformers (BERT). Deep learning requires a large number of labeled data to work. Meanwhile, deep learning often implements the semi-supervised learning (SSL) method, which is able to use massive unlabeled data to improve performance of supervised deep learning. As an effective SSL variant, unsupervised data augmentation (UDA) focuses on data augmentation techniques to improve the performance of deep learning. We present an enhanced version of UDA (EUDA) by mixing more data augmentation strategies and using a problem related prefilter. Our EUDA targets at emergency event detection and classification. Considering that emergency events always have time and location elements, text can be filtered based on this semantic feature. We propose to use semantic feature aided enhanced unsupervised data augmentation to solve the concerned problem. Empirical studies on the dataset prepared for the task validates that the proposed EUDA can achieve significantly better performance than supervised learning with a limited size of labeled data. Experiments are also carried out on a text classification task, which confirms that EUDA improves performance for BERT neural network.