Deep Learning-Based Sentiment and Stance Analysis of Tweets About Vaccination

Introduction

Automatic analysis of social media posts to extract valuable pieces of information is increasingly employed in different domains. Significant applications of social media analysis on Twitter include information extraction from tweets, tweet summarization, and analysis of tweets to determine users’ sentiments and stances based on their tweets. Such applications are proposed and described for domains including public health, education, law, energy, and finance.

Concerning public health, tweets can be automatically analyzed in order to protect and improve public health and to help decision-makers (of public health) by providing them with the analysis results. Several studies outline the possible contributions of such results to the public health domain (Culotta, 2014; Zhou et al., 2018; Küçük et al., 2021).

As a recent example, COVID-19 pandemic has been an important public health problem, and it is emphasized in related studies that people express their opinions about the pandemic and its many aspects through social media channels like Twitter, during the period of the pandemic (Burzyńska et al., 2020; Alkhaldi et al., 2022; Mohammed et al., 2022). Hence, related studies aim to explore Twitter users’ opinions about the pandemic and other related aspects, such as using face masks, school closures, remote working, lockdowns, and COVID-19 vaccines.

In this paper, we focus on applying deep learning-based sentiment analysis and stance detection on Turkish tweets concerning COVID-19 vaccination/vaccines. During sentiment analysis, individual tweets’ polarity (or valence) is considered with the related classes of positive, negative, and none, while during stance detection, COVID-19 vaccination is taken as the target of stance, and the stances in tweets towards this target are investigated using the classes of favor, against, and none. The performance of a transformer-based deep learning model (i.e., BERT) for both problems is compared with the results of (traditional) SVM and random forest models. Additionally, we test the performance of another popular deep learning model, ChatGPT¹, for both tasks on our dataset, and compare its performance with the aforementioned BERT model. Both BERT and ChatGPT are transformer-based pre-trained language models and, to the best of our knowledge, this study is the first one to employ ChatGPT for both tasks in this application settings and also the first study to compare the performance of BERT and ChatGPT for the analysis of health-related social media.

The main contributions of the current paper can be listed as follows:

• •

  This paper showcases joint application of two significant research tasks of affective computing in the public health domain. Both sentiment analysis and stance detection are significant tasks that output complementing results, from which health professionals can readily benefit during their decision and policy making processes.

• •

  The COVID-19 pandemic has been a long, challenging, and tiresome period for all of the individuals in the world. Fortunately, COVID-19 vaccines effectively help end this pandemic and alleviate its effects. Therefore, it is of utmost importance (for health-related decision-makers) to discover the opinions and stances of the public towards COVID-19 vaccination/vaccines. Using the findings of the current study, precautions can be taken to combat vaccine hesitancy for COVID-19 and other infectious diseases in society.

• •

  In addition to applying the BERT model for stance and sentiment detection on the tweet dataset about vaccination, we have also tested the performance of ChatGPT for both tasks on the same dataset.

• •

  This paper also includes a system proposal for a large-scale public health surveillance and decision support system encompassing the sentiment analysis and stance detection modules described in the paper.

• •

  A significant review of the related recent literature is also included in the current study, which will contribute to researchers who aim to work on social media analysis for public health monitoring.