The impact of information along with the innovations in the development of web platforms have enabled people across the globe to connect and share content in many different ways. This change in communication and human interaction has resulted in the generation of great amounts of textual data. Social Networks, web fora, chat rooms, product reviews reflect the human need to communicate, argue for or against a subject. These large volumes of data stored in the cyber space pose a lot of challenges to the existing algorithms of knowledge discovery from data. User generated data if extracted and analyzed appropriately can provide valuable information and help us understand markets, political conflicts or social behaviour. On the other hand humans are unable to grasp, let alone analyze, such a huge amount of data. Thus, machine learning techniques and models must be designed and utilized to make sense of these very large data sets.
Such models would be useful to the average user if they were able to track, extract and exploit these data sets. Moreover, such models should also be able to assess the learned knowledge or experience by sensing and adapting to background changes as these are manifested from the variability of user's opinions. This PhD thesis advances the state-of-the-art in opinion mining by a) outperforming the generalization of state-of-the-art sentiment analysis methods, through a sentiment analysis method that bridges two sentiment assigning processes (chapter 3), b) mining explicit and implicit aspects in opinionated text in a more efficient way due to an innovative model development that combines two different types of aspect extraction encodings (chapter 4), c) providing algorithms that are able to extract domain specific opinion words for dictionary-based sentiment analysis (chapter 5), d) we have designed and fine-tuned our algorithms for applications of opinion mining in interdisciplinary areas like crowd-sourcing, political science, and
conversational agents (chapter 5).