Sentiment analysis of social media data using Supervised learning and Ekman’s basic emotions theory

Abstract

Social media has emerged as an effective source to investigate people’s opinions in the context of a variety of topics and situations, including crime. Crime solving can be a difficult task hence requiring human intelligence together with experience. Crime data is usually big and full of noise; hence manual analysis of this data is tedious and sometimes impossible to do. Data mining techniques such as Sentiment Analysis can help in analysing such big data. In Sentiment Analysis, polarity in a text is identified using text processing and classification. For our study, we used textual sentiment analysis in retrieving the sentiment or emotion carried by a piece of text at sentence level. We used multi-classification of data by categorising data using the Ekman’s basic emotions theory. This research work applied Sentiment Analysis of Facebook data in order to predict the public emotion regarding crime issues affecting their lives. A model was designed and trained using supervised learning approaches; Naïve Bayes, J48, K-NN and Random Forest. The model was tested on Facebook crime dataset. The results from the experiments showed that Random Forest outperformed all the other classifiers with an accuracy of 80%. Naïve Bayes followed with an accuracy of 71.29%, K-NN with 69.21% and J48, which was the least performing classifier, achieved 65.92%. The results of the predictive model were also used to demonstrate a correlation between the moods observed on social media and the Botswana Police Annual Report of 2016. Results show that sentiments have an impact on the outcome of the Police Report, hence showing a positive correlation. Results provide valued information that will assist the Botswana Police Service to put in strategy their actions with considerations to public sentiments and hence improve the process of handling crime related issues.