A Review on Sentiment Analysis from Social Media Platforms

Abstract

Sentiment analysis has proven to be a valuable tool to gauge public opinion in different disciplines. It has been successfully employed in financial market prediction, health issues, customer analytics, commercial valuation assessment, brand marketing, politics, crime prediction, and emergency management. Many of the published studies have focused on sentiment analysis of Twitter messages, mainly because a large and diverse population expresses opinions about almost any topic daily on this platform. This paper proposes a comprehensive review of the multifaceted reality of sentiment analysis in social networks. We not only review the existing methods for sentiment analysis in social networks from an academic perspective, but also explore new aspects such as temporal dynamics, causal relationships, and applications in industry. We also study domains where these techniques have been applied, and discuss the practical applicability of emerging Artificial Intelligence methods. This paper emphasizes the importance of temporal characterization and causal effects in sentiment analysis in social networks, and explores their applications in different contexts such as stock market value, politics, and cyberbullying in educational centers. A strong interest from industry in this discipline can be inferred by the intense activity we observe in the field of intellectual protection, with more than 8,000 patents issued on the topic in only five years. This interest compares positively with the effort from academia, with more than 2,300 articles published in 15 years. But these papers are unevenly split across domains: there is a strong presence in marketing, politics, economics, and health, but less activity in other domains such as emergencies. Regarding the techniques employed, traditional techniques such as dictionaries, neural networks, or Support Vector Machines are widely represented. In contrast, we could still not find a comparable representation of advanced state-of-theart techniques such as Transformers-based systems like BERT, T5, T0++, or GPT-2/3. This reality is consistent with the results found by the authors of this work, where computationally expensive tools such as GPT-3 are challenging to apply to achieve competitive results compared to those from simpler, lighter and more conventional techniques. These results, together with the interest shown by industry and academia, suggest that there is still ample room for research opportunities on domains, techniques and practical applications, and we expect to keep observing a sustained cadence in the number of published papers, patents and commercial tools made available.