Using Artificial Intelligence to Understand What Causes Sentiment Changes on Social Media

Fuad Alattar,Khaled Shaalan

doi:10.1109/access.2021.3073657

Abstract

Sentiment Analysis tools allow decision-makers to monitor changes of opinions on social media towards entities, events, products, solutions, and services. These tools provide dashboards for tracking positive, negative, and neutral sentiments for platforms like Twitter where millions of users express their opinions on various topics. However, so far, these tools do not automatically extract reasons for sentiment variations, and that makes it difficult to conclude necessary actions by decision-makers. In this paper, we first compare performance of various Sentiment Analysis classifiers for short texts to select the top performer. Then we present a Filtered-LDA framework that significantly outperformed existing methods of interpreting sentiment variations on Twitter. The framework utilizes cascaded LDA Models with multiple settings of hyperparameters to capture candidate reasons that cause sentiment changes. Then it applies a filter to remove tweets that discuss old topics, followed by a Topic Model with a high Coherence Score to extract Emerging Topics that are interpretable by a human. Finally, a novel Twitter’s sentiment reasoning dashboard is introduced to display the most representative tweet for each candidate reason.

Highlights

Hundreds of millions of tweets are being posted every day to discuss various topics [1] like politics, products, news, celebrities, etc
We focus on the problem of automatic discovery of reasons behind sentiment variations on social media
2) Foreground-Background Latent Dirichlet Allocation (FB-LDA) FB-LDA Model was developed by Tan et al [6] who manually analyzed real-life tweets on certain targets and noticed that main reasons for sentiment variations are causally linked to Emerging Topics

Summary

INTRODUCTION

Hundreds of millions of tweets are being posted every day to discuss various topics [1] like politics, products, news, celebrities, etc. Step includes applying LDA on the tweets of the spike to identify the topic that has highest frequency as it is assumed this topic is the main reason for the sentiment spike. FILTERED-LDA FOR SENTIMENT REASONING Inspired by the FB-LDA Model, we introduce a Filtered-LDA framework, which aims to overcome the four main limitations of FB-LDA by (1) Enhancing the topic categorization accuracy through ensuring low Perplexity Score for the model and applying multiple settings of LDA hyperparameters to perform a deep scan for discussed topics, (2) removing all documents that include old/background topics to ensure that final output will include Emerging Topics only, (3) enhance the interpretability of detected Emerging Topics by using the highest LDA Coherence Score and reducing the chance of using words from old/background topics, and (4) use accurate sentiment variation criteria. It shows the highest frequency topics which were discussed on each day of the Ground Truth Dataset, wherein the SMS Vulnerability topic is the Emerging Topic, which caused major sentiment variation

COMPARING SENTIMENT ANALYSIS CLASSIFIERS

FINDING REASON CANDIDATES

Findings

CONCLUSION