From Conditional Random Field (CRF) to Rhetorical Structure Theory (RST): incorporating context information in sentiment analysis

This paper investigates a method based on Conditional Random Fields (CRFs) to incorporate sentence structure (syntax and semantics) and context information to identify sentiments of sentences. It also demonstrates the usefulness of the Rhetorical Structure Theory (RST) taking into consideration the discourse role of text segments. Thus, this paper’s aim is to reconsider the effectiveness of CRF and RST methods in incorporating the contextual information into Sentiment Analysis systems. Both methods are evaluated on two, different in size and genre of information sources, the Movie Review Dataset and the Finegrained Sentiment Dataset (FSD). Finally, we discuss the lessons learned from these experimental settings w.r.t. addressing the following key research questions such as whether there is an appropriate type of social media repository to incorporate contextual information, whether extending the pool of the selected features could improve context incorporation into SA systems and which is the best performing feature combination to achieve such improved performance