Surviving in the Social Media Jungle: sentiment Analysis through Structured Learning Paradigms

The analysis of emotions is a cross-disciplinary research field, with theoretical underpinnings in philosophy, psychology, linguistics and computer science. Emotions are an important element of the human nature as they are complex states of feeling that have influence in our physical and psychological reactions. Emotions, influence cognition and intelligence, especially when this involves social decision-making and interaction. For these reasons, the analysis of emotions is a crucial step in making computers able to interact with humans more naturally. The study of emotions is based on the observation of emotional states starting from natural language or body expressions. The aim of Natural Language Processing, Opinion Mining and Sentiment Analysis is to analyze emotions and sentiment from texts. In recent years, a central role is embodied by the analysis of Social Networks and Social Media as people express their feelings and thoughts preferably through these channels. Sentiment Analysis in Social Media has been in general tackled with Machine Learning, by treating the sentiment recognition problem as Text Classification task over single instances. In this Thesis, we design models for Sentiment Analysis that are sensitive to the di↵erent aspects that can characterize the expression of attitudes of a person in a text. Support Vector Machine classifiers along with Kernel Methods allow us to define and properly combine isolated and apparently non-correlated dimensions of the sentiment, e.g. sentiment bearing words or ironic characteristics. We will provide methodologies to derive such observations directly from data that do not need human supervision. The properties we are going to define are general enough and mainly language independent: it makes the proposed solutions appealing from an engineering perspective. Often people write their messages by reusing as much as possibile external knowledge embodied, for example, in v already written messages. The analysis of the sentiment in Social Media cannot leave aside this rich set of information, as often it is the only useful data that can discriminate between opposite polarity messages. In this Thesis, we are going to immerse messages in their “contexts”, as this information is important in correctly recognizing an emotion, as advocated by some scientists [Pantic(2009)]. We propose the study of sentiment phenomena by making computational learning models aware of the context in which a message lives. A proper Structured Learning framework, based on a Markovian formulation of Support Vector Machine, will be discussed and adopted to make the classification of a single short message dependent on the available contextual information. This framework is intended to capture linear relationships between messages, by considering sequences of them if related by some structural or topical characteristic.