Analysing social media data using sentiment analysis in relation to public order

Abstract

The research aim is to analyse social media data using sentiment analysis in relation to public order. A sentiment can be expressed in a thought, opinion or attitude that is mainly based on emotion instead of reason. (SA) Sentiment Analysis studies the opinions, sentiments and emotions expressed at sentence or document level. SA extracts text which is identified and classified as opinions or emotions that aim to support a decision-making process through the analysis of text. SA identifies and measures whether the text being analysed is positive, negative or neutral in relation to an entity, such as people, organisation, event, location, or a topic. As the adoption of ubiquitous technology increases and the population on social media continues to grow with the speed of responsiveness of the users expressing their political, economic or religious views on Twitter or Facebook, the posts become valuable sources of public opinion. This can be seen as an important commodity to be used to infer public opinions for social studies or marketing. The research suggests the police have found it difficult to adapt their existing model to the changing nature of public events and handling of acceleration towards technology and social media. The scalability and volume of data has made it increasingly hard for the police to manage, monitor and make use of intelligence emerging from social media to maintain the peace. To address this gap, the investigation will evaluate whether SA can enhance the analysis of social media in the context of public (dis)order events. This may help to improve the police’s decision-making process and reduce complexity to increase public safety. There are specific and generalised ways that SA can support the police, but this research might focus on a specific case. To meet the aim, the research proposes to use a SA model, data mining tools and techniques to analyse the relevant data extracted from social media. The project will use an adapted social media lifecycle as a methodological approach. Past events involving public order and the police will be evaluated to develop relevant methodology and provide appropriate recommendations to the technical community on ways to use SA for future applications of social media. In the project it adopted a hybrid approach which consists of a dictionary, machine learning and gold standard approaches. As result, the machine learning of dictionaries and manual classification results proved to show the strongest output based on precision, recall and F1 measure when compared to the machine learning of tweets and manual classification. The change point analysis helped to identify significant points in the timeline of tweets for the event which correlated to the physical event. However, there were some inaccuracies on the allocated points of change, as deemed insignificant based on news media and low volume of tweets. Future work is required to understand the reasons behind the allocation change points and possible use of alternative methods to help extract further insights that could not be explored in this project. The study makes a series of contributions to knowledge. First, to the creation of a keywords for public order events due to none being publicly available. Second, is to build towards a model to predict what may happen in public order events with the application of dictionary, machine learning and creation of gold standard in the realm of sentiment analysis. Third, the technical contribution to sentiment analysis community to help provide future recommendations to potentially enhance their framework and what areas require further research in the area. Fourth, is the development of social media lifecycle methodology, which has been tested in this project.

Comparative Analysis of Lexicon-Based Sentiment Analysis Methods

“…Sentiment analysis can be applied to various tasks, but we focus on comparing polarity of short text on a sentence-level [2]. The detection of polarity is common across sentiment methods that provides important insights to a series of different applications, social media is one that can be commonly sourced.…”

“…A limitation of supervised method is the need of labelled data, which can be resource intensive. A lexical based approach is where has a pre-defined list of words, where each word is assigned a polarity score, but the lexical method may vary on their output dependent on the context on how they were created [2]. For instance, VADAR lexicon was to discover patterns of smoking and drinking abstinence within social media data [8].…”

“…There are many dictionaries (e.g., SentiWordNet, SenticNet, Stanford and SentiStrength) that are based on the English language, but most are American English rather than UK English [1]. Furthermore, other dictionaries with different languages are sparse in comparison with English based dictionaries [2]. These dictionaries may have a wider term coverage, but there are a comparatively limited number of words with a fixed sentiment orientation, or score assigned to the words [2].…”

“…Furthermore, other dictionaries with different languages are sparse in comparison with English based dictionaries [2]. These dictionaries may have a wider term coverage, but there are a comparatively limited number of words with a fixed sentiment orientation, or score assigned to the words [2].…”

“…Our aim is to classify tweets as positive, negative and neutral, but to do this in a highly automated way to adapt to the high volume of social media data. We have chosen to adopt the hybrid approach, combining both lexicon and machine learning approaches to apply sentiment analysis to four Twitter datasets [2]. The lexicon-based approach will perform at sentence level to determine the polarity from the predefined dictionary while the machine learning algorithms including Support Vector Machine (SVM), Na¨ıve Bayes and Maximum Entropy will train a classifier by using the polarity for each sentence as determined by the lexicon [2].…”