Supervised and semi-supervised statistical models for word-based sentiment analysis

Ever since its inception, sentiment analysis has relied heavily on methods that use words as their basic unit. Even today, such methods deliver top performance. This way of representing data for sentiment analysis is known as the clue model. It offers practical advantages over more sophisticated approaches: It is easy to implement and statistical models can be trained efficiently even on large datasets. However, the clue model also has notable shortcomings. First, clues are highly redundant across examples, and thus training based on annotated data is potentially inefficient. Second, clues are treated context-insensitively, i.e., the sentiment expressed by a clue is assumed to be the same regardless of context. In this thesis, we address these shortcomings. We propose two approaches to reduce redundancy: First, we use active learning, a method for automatic data selection guided by the statistical model to be trained. We show that active learning can speed up the training process for document classification significantly, reducing clue redundancy. Second, we present a graphbased approach that uses annotated clue types rather than annotated documents which contain clue instances. We show that using a random-walk model, we can train a highly accurate document classifier. We next investigate the context-dependency of clues. We first introduce sentiment relevance, a novel concept that aims at identifying content that contributes to the overall sentiment of the review. We show that even when we have no annotated sentiment relevance data available, a high-accuracy sentiment relevance classifier can be trained using transfer learning and distant supervision. Second, we perform linguistically motivated analysis and simplification of a compositional sentiment analysis. We find that the model captures linguistic structures poorly. Further, it can be simplified without any loss of accuracy.