DFDS: A Domain-Independent Framework for Document-Level Sentiment Analysis Based on RST

Document-level sentiment analysis is among the most popular research fields of nature language processing in recent years, in which one of major challenges is that discourse structural information can be hardly captured by existing approaches. In this paper, a domain-independent framework for document-level sentiment classification with weighting rules based on Rhetorical Structure Theory is proposed. First, original textual documents are parsed into rhetorical structure trees through a preprocessing pipeline. Next, the sentiment score of elementary discourse units is computed via sentence-level sentiment classification method. Finally, according to the rhetorical relation between neighbor discourse units, we define weighting schema and composing rules based on which scores of elementary discourse units are summed recursively to the whole document. Experiment results show that our approach has better performance on datasets in different domains, compared with state-of-art document-level sentiment analysis systems based on RST, and the best result is 15% higher than baseline.