Political Opinion Mining and Analysis for Social Media

Computational approaches to opinion mining have mostly focused on sentiment classification for reviews or blogs such as product reviews. This thesis analyses socio-political opinion mining to determine sentiment or stance towards a given target of interest, particularly for social media such as Twitter and Facebook. We explore the tasks of automatically determining from political opinion in social media whether the author of a given text is positive or negative towards a target and aspects of the target, a problem known as fine-grained opinion mining, and in favour or against a target, a problem known as stance detection. The main contribution of this thesis is the introduction of two novel probabilistic topic models for fine-grained opinion mining and stance detection for political opinion analysis from social media. Such models are capable of representing all the variables of interest simultaneously and considering dependencies between variables. Further, the ability to train the models in a weakly supervised approach, without the use of labelled corpora, is another advantage of the proposed models. The first model, called Joint Entity-Aspect-Sentiment (JEAS), is based on Latent Dirichlet Allocation (LDA), which detects the target entity, aspects of the entity and sentiment towards entity and aspect simultaneously from the text. The underlying hypothesis of the JEAS is that sentiment of an opinion is dependent on both target entity and aspect. The second model, Joint Sentiment-Issue-Stance (JSIS), also based on LDA, detects target issues and stance simultaneously from text. In JSIS, the stance is generated conditioned on both issues and sentiment distributions. The proposed models have been evaluated on Twitter and Facebook datasets in the political domain. The experimental results of both models illustrate the effectiveness of the proposed models and confirm the validity of the underlying hypotheses of each model by outperforming baselines for document-level classification tasks. Additionally, both models are capable of learning coherent and informative topic words related to each variable in the models