Crestani, Fabio Giachanou, Anastasia 2018-10-16 The increasing popularity of social media has changed the web from a static repository of information into a dynamic forum with continuously changing information. Social media platforms has given the capability to people expressing and sharing their thoughts and opinions on the web in a very simple way. The so-called User Generated Content is a good source of users opinion and mining it can be very useful for a wide variety of applications that require understanding the public opinion about a concept. For example, enterprises can capture the negative or positive opinions of customers about their services or products and improve their quality accordingly. The dynamic nature of social media with the constantly changing vocabulary, makes developing tools that can automatically track public opinion a challenge. To help users better understand public opinion towards an entity or a topic, it is important to: a) find the related documents and the sentiment polarity expressed in them; b) identify the important time intervals where there is a change in the opinion; c) identify the causes of the opinion change; d) estimate the number of people that have a certain opinion about the entity; and e) measure the impact of public opinion towards the entity. In this thesis we focus on the problem of tracking public opinion on social media and we propose and develop methods to address the different subproblems. First, we analyse the topical distribution of tweets to determine the number of topics that are discussed in a single tweet. Next, we propose a topic specific stylistic method to retrieve tweets that are relevant to a topic and also express opinion about it. Then, we explore the effectiveness of time series methodologies to track and forecast the evolution of sentiment towards a specific topic over time. In addition, we propose the LDA & KL-divergence approach to extract and rank the likely causes of sentiment spikes. We create a test collection that can be used to evaluate methodologies in ranking the likely reasons of sentiment spikes. To estimate the number of people that have a certain opinion about an entity, we propose an approach that uses pre-publication and post- publication features extracted from news posts and users' comments respectively. Finally, we propose an approach that propagates sentiment signals to measure the impact of public opinion towards the entity's reputation. We evaluate our proposed methods on standard evaluation collections and provide evidence that the proposed methods improve the performance of the state-of-the-art approaches on tracking public opinion on social media. application/pdf https://susi.usi.ch/global/documents/319055 https://n2t.net/ark:/12658/srd1319055 https://susi.usi.ch/documents/319055/files/2018INFO016.pdf eng info:eu-repo/semantics/altIdentifier/urn/urn:nbn:ch:rero-006-117948 info:eu-repo/semantics/altIdentifier/ark/12658/srd1319055 info:eu-repo/semantics/openAccess License undefined Opinion dynamics Opinion mining Opinion retrieval Sentiment analysis Tracking sentiment Sentiment spikes Emotional reactions Reputation analysis Time series Social media info:eu-repo/classification/udc/004 Tracking public opinion on social media http://purl.org/coar/resource_type/c_db06