Tracking Nicotine and COVID-19 Discourse on Quora : Infodemiology Study Using Topic Modeling and Sentiment Analysis (Preprint)

Abstract

Misinformation in public forums like Quora, regarding vaping or smoking nicotine being a therapeutic and/or preventive treatment for COVID 19 has potential to undermine the efforts to mitigate the pandemic and can be detrimental to public health.

To identify sentiments and topics in the COVID-19 and nicotine related discourse on Quora between March 11, 2020, and August 31, 2021, to better understand public perceptions of the misinformation regarding vaping or smoking nicotine being a therapeutic and/or preventive treatment for COVID-19.

Data was obtained from Quora by keyword search followed by web scrapping using Python’s BeautifulSoup library. Topic modelling using Latent Dirichlet Allocation (LDA) model was used to understand underlying semantics and word prevalence within selected archetypes for the study. Sentimental analysis was performed using hybrid approach that involved two steps: (i) lexicon-based approach with SentiStrength tool was used to capture the positive, negative, and neutral sentiments in the data from a pre-defined lexicon dictionary; (ii) machine learning (ML) approach was used to train the polarities obtained from Step 1 using the Sci-kit (sk) learn framework. The hybrid approach to sentiment analysis resulted in higher accuracy and higher precision of the results.

A total of 19,252 documents were obtained for the analyses. 12 distinct word clouds were generated from the Quora data. Based on their word distribution, 11 clouds were assigned distinct archetypes related to the research objective and one cloud could not be labelled. The results of the sentimental analyses revealed that majority (50%) of the population exhibited neutral sentiments, followed by 26%, that exhibited positive sentiments. Only 24% of the population have negative sentiments against this statement. The XGBOOST algorithm was applied on the data by splitting it into 80% training and 20% testing to improve the performance measurements. Performance evaluation resulted in an accuracy of 86% after fine tuning of the dataset. XGBOOST was the best preforming ML model for this study.

This research can aid the public health officials to craft the right messages and right education tools for the safety and benefit of the public during such pandemics.