Considering Uncertainty Expression in Sentiment Analysis and Tweet Classification

Abstract

Expressing uncertainty on social media differs from formal language, allowing authors to write in their preferred styles. Investigative activities face challenges in recognizing the level of confidence conveyed in social media texts. Despite existing corpora in other domains, limited attention has been given to the aspect of uncertainty in microblogging. This research focuses on analyzing sentiments expressed on Twitter while considering the semantic uncertainties present within tweets, particularly in relation to the Covid-19 pandemic. A tweet classification algorithm is developed to assess uncertainty and sentiment. The tweets are categorized as “certain” or “uncertain,” with further subcategories of uncertainty including “question,” “condition,” “hope,” and “belief.” The performance obtained from the algorithm demonstrates its effectiveness. Our study found uncertainty in around one-third of tweets, primarily in the form of questions. With regard to sentiments, neutrality was dominant, followed by positivity, while the belief category leaned toward positivity. The research highlights the significance of recognizing uncertainty on social media using contextual semantic cues rather than traditional indicators. Additionally, exploring sub-classes of uncertainty provides valuable insights for managing uncertainty in social media texts. Careful consideration of relevant semantic categories in sentiment analysis, excluding biased categories, is crucial. Based on the findings, refining the “belief” category by considering nuanced types, such as doubt, hesitation, and presumption, is recommended. This refinement would benefit domains focused on truth discovery and investigation. Furthermore, studying the correlation between uncertainty expression and the truth value of statements is suggested, providing deeper insights into how uncertainty influences credibility and truthfulness.