Topic Sentiment Trend Detection and Prediction for Social Media

Abstract

Social media often plays a crucial role in disseminating information to warn the public about health concerns. Opioid addiction has become of the significant outbreaks in the United States. Studying opioid issues in social media has the potential to reveal patterns of opioid abuse and understand people's opinions on this issue. On the other hand, social media forums like Twitter allow for open discussions among the public and popular information exchanges. The question arises if the trends of such concerns like opioid abuse can be automatically detected and predicted for a better understanding of people's attitude changes in a specific direction.

In this thesis, we developed a novel framework for topic sentiment trend detection and prediction in social media. The proposed framework was designed to cope with the following tasks: topic trend detection, sentiment analysis, and topic prediction. The VADER-based time series sentiment analysis and the KATE-based topic modeling methods were applied to analyze the social media data from the public, social media news, and newspapers. We have further extended the framework for the successful prediction of topic trends for given the current issues. For the topic trend prediction model, the deep neural network model called Long Short Term Memory (LSTM) was used along with topic embedding techniques. The two-step communication model was used to evaluate how different types of media are active in the emergence and escalation dissemination and how effective it is in the pacification and prevention of concerns related to epidemics, like opioids. The proposed framework was also applied to the Twitter data as well as New York Times articles on opioids over ten years, from 2010 to 2019. The results of this study have shown some exciting findings from topic sentiment detection and high accuracies from the topic trend prediction.