Neural Networks and Support Vector Machine based Approach for Classifying Tweets by Information Types at TREC 2018 Incident Streams Task

Microblog, especially twitter, is treated as an important source to serve the situational information needs during a disaster period. Monitoring and producing the curated tweets based on different information types from massive twitter posts provide enormous opportunities to different public safety personnel or used for post-incident analysis. In this paper, we present our approach to addressing the problem defined in the TREC 2018 incident streams (TREC-IS) task. The task is to classify the tweets in each event/incident’s stream into different high-level information types within the incident ontology. In our approach, we employ different deep neural network (DNN) classifiers in combination with a multi-class support vector machine (SVM) classifier and a rule-based classifier. We consider a rich set of hand-crafted features to train our multi-class SVM classifier, whereas a pre-trained word2vec model is used for the DNN based classifiers. Moreover, we introduce a set of rules based on the language of tweets, exploiting indicator terms, and WH-orientation of tweets for our rule-based classifier. Experimental results showed that our proposed KDEIS4 DM method obtained the second position among the participants in TREC-IS task and outperforms the participant median by more than 8% and 5% in terms of F1 Score and accuracy, respectively.