Abstract
Automated sentiment analysis has become an active field of study with a broad applicability. One of the key open research issues lies in dealing with structural aspects of text when analyzing its conveyed sentiment. Recent work uses structural aspects of text in order to distinguish important text segments from less important ones in terms of their contribution to the overall sentiment. Yet, existing methods are conned to making coarse-grained distinctions between text segments based on segments' rhetorical roles, while not accounting for the full hierarchical rhetorical structure in which these roles are dened. We hypothesize that a better understanding of a text's conveyed sentiment can be obtained by guiding automated sentiment analysis by the full rhetorical structure of text. We evaluate our hypothesis in a framework for sentiment analysis that is based on Rhetorical Structure Theory, at the level of sentences, paragraphs, and documents. On an English movie review corpus, we obtain signicant classication performance improvements compared to baselines not or only shallowly accounting for rhetorical structure, with the best results generated by exploiting a text's full sentential rhetorical structure.
Original language | English |
---|---|
Pages (from-to) | 69-77 |
Number of pages | 9 |
Journal | Communications of the ACM |
Volume | 58 |
Issue number | 7 |
DOIs | |
Publication status | Published - Jul 2015 |
Keywords
- 2024 OA procedure