Discourse-level Relations For Opinion Analysis

Opinion analysis deals with subjective phenomena such as judgments, evaluations, feelings, emotions, beliefs and stances. The availability of public opinion over the Internet and face to face conversations; coupled with the need to understand and mine these for end applications has motivated a great amount of research in this field in recent times. Researchers have explored a wide array of knowledge resources for opinion analysis, from words and phrases to syntactic dependencies and semantic relations.In this thesis, we investigate a discourse-level treatment for opinion analysis.In order to realize the discourse-level analysis, we propose a new linguistic representational scheme designed to support interdependent interpretations of opinions in the discourse. We adapt and extend an existing subjectivity annotation scheme to capture discourse-level relations in multi-party meeting corpus. Human inter-annotator agreement studies show that trained human annotators can recognize the elements of our linguistic scheme. Empirically, we test the impact of our discourse-level relations on fine-grained polarity classification. In this process, we also explore two different global inference models for incorporating discourse-based information to augment word-based information. Our results show that the discourse-level relations can augment and improve upon word-based methods for effective fine-grained opinion polarity classification. Further, in this thesis, we explore linguistically motivated features and a global inference paradigm for learning the discourse-level relations form the annotated data. We employ the ideas from our linguistic scheme for recognizing stances in dual-sided debates from the product and political domains. For product debates, we use web mining and rules to learn and employ elements of our discourse-level relations in an unsupervised fashion. For political debates, on the other hand, we take a more exploratory, supervised approach, and encode the building blocks of our discourse-level relations as features for stance classification. Our results show that, the ideas behind the discourse level relations can be learnt and employed effectively to improve overall stance recognition in product debates