SVM-Based Domain Adaptation Machine Learned Models for the Automatic Classification of Disaster-Related Tweets

Abstract

Abstract-Twitter is a well-known microblogging service used by many users to quickly acquire disaster-related information. Since there is an influx of tweet during the occurrence of a disaster event, the need to automatically filter informative disaster-related tweets is imperative. In this study, the researchers used the Habagat tweets and tweets gathered during the Zamboanga Siege event as training corpora for training two Support Vector Machine (SVM) machine classifiers. The first machine classifier is used to filter informative tweets while the second model is used to classify informative tweets as Donation, Traffic update, Weather update, Class suspension, Rescue and Relief and River level/condition. These subclasses were based on the latent topics extracted from the corpus of informative tweets using Principal Component Analysis(PCA) and Latent Dirichlet Allocation(LDA). Several experiments using different preprocessing techniques, vector space of unigrams/bigrams with TF-IDF weightings on balanced and unbalanced data were applied to test and train the classifiers. SVM and Expectation Maximization (EM) algorithms were used for the training and testing of the machine classifiers. The models were evaluated based on the performance metrics of accuracy, precision and recall. Experimental results show that for a binary classifier, using unigram with a corpus of unbalanced labeled corpus returned best classification results and applying bigram to a balanced corpus gave high performance results. Furthermore, the multiclass model yielded better performance results with unigrams as features.