Corpus-based analysis of rhetorical relations: A study of lexical cues

In spite of the long tradition of Rhetorical Structure Theory (RST) in computational linguistics, there is no robust method capable of detecting rhetorical relations in the text of discourse. To pave the way for development of such techniques, we carried out experiments aimed at understanding the effectiveness of using corpus-based lexical cues in the identification of RST relations for three different relations and across two different text genres. In particular, we focused on the three relations of CIRCUMSTANCE, EVALUATION, and ELABORATION and two different corpora: newspaper articles and online reviews. The analysis results indicate that the cue-based approaches can be quite effective in detecting CIRCUMSTANCE. However, the ability of lexical cues in relation identification is limited for ELABORATION. For the EVALUATION relation, genre-specific factors can play a more significant role.