Twitter sentiment classification for measuring public health concerns

An important task of public health officials is to keep track of health issues, such as spreading epidemics. In this paper, we are addressing the issue of spreading public concernabout epidemics. Public concern about a communicable disease can be seen as a problem of its own. Keeping track of trends in concern about public health and identifying peaks of public concern are therefore crucial tasks. However, monitoring public health concerns is not only expensive with traditional surveillance systems, but also suffers from limited coverage and significant delays. To address these problems, we are using Twitter messages, which are available free of cost, are generated world-wide, and are posted in real time. We are measuring public concern using a two-step sentiment classification approach. In the first step, we distinguish Personal tweets from News (i.e., Non-Personal) tweets. In the second step, we further separate Personal Negative from Personal Non-Negative tweets. Both these steps consist themselves of two sub-steps. In the first sub-step (of both steps), our programs automatically generate training data using an emotion-oriented, clue-based method. In the second sub-step, we are training and testing three different Machine Learning (ML) models with the training data from the first sub-step; this allows us to determine the best ML model for different datasets. Furthermore, we are testing the already trained ML models with a human annotated, disjoint dataset. Based on the number of tweets classified as Personal Negative, we compute a Measure of Concern (MOC) and a timeline of the MOC. We attempt to correlate peaks of the MOC timeline to peaks of the News (Non-Personal) timeline. Our best accuracy results are achieved using the two-step method with a Naive Bayes classifier for the Epidemic domain (six datasets) and the Mental Health domain (three datasets).