Analysing Event-Related Sentiments on Social Media with Neural Networks

Abstract

Sentiment analysis is performed to determine the polarity of opinion on a subject. It has been applied to text corpora such as movie reviews, financial documents to glean information about overall-sentiment anc produce actionable data. Recent events have demonstrated that polling can be sometimes unreliable. People can be difficult to access through conventional polling methods and less than frank in polls. In the era of social media, voters are likely to more freely express their opinion on social media forums about divisive events especially in media where anonymity exists. Analyzing the prevailing opinion on these forums can indicate if there are any deficiencies in polling and can be a valuable addition to conventional polling. We analyzed text corpora from Reddit forums discussing the recent referendum in Britain to exit from the EU (known as Brexit). Brexit was an important world event and was very divisive in the run-up and post vote. We analyzed sentiment in two ways: Initially we tried to gauge positive, negative, and neutral sentiments. In the second analysis, we further split these sentiments into six different polarities based on the directionality of the positive and negative sentiments (for or against Brexit). Our technique utlilized paragraph vectors (Doc2Vec) to construct feature vectors for sentiment analysis with a Multilayer Perceptron classifier. We found that the second analysis yielded overall better results; although, our classifier didn’t perform as well in classifying positive sentiments. We demonstrate that it is possible glean valuable information from complicated and diverse corpora such as multi-paragraph comments from reddit with sentiment analysis.