Improving Sentiment Analysis with Document-Level Semantic Relationships from Rhetoric Discourse Structures

Conventional sentiment analysis usually neglects semantic information between (sub-)clauses, as it merely implements so-called bag-of-words approaches, where the sentiment of individual words is aggregated independently of the document structure. Instead, we advance sentiment analysis by the use of rhetoric structure theory (RST), which provides a hierarchical representation of texts at document level. For this purpose, texts are split into elementary discourse units (EDU). These EDUs span a hierarchical structure in the form of a binary tree, where the branches are labeled according to their semantic discourse. Accordingly, this paper proposes a novel combination of weighting and grid search to aggregate sentiment scores from the RST tree, as well as feature engineering for machine learning. We apply our algorithms to the especially hard task of predicting stock returns subsequent to financial disclosures. As a result, machine learning improves the balanced accuracy by 8.6 percent compared to the baseline.