Automatic identification of rhetorical relations among intra-sentence discourse segments in Arabic

Identifying discourse relations, whether implicit or explicit, has seen renewed interest and remains an open challenge. We present the first model that automatically identifies both explicit and implicit rhetorical relations among intra-sentence discourse segments in Arabic text. We build a large discourse annotated corpora following the rhetorical structure theory framework. Our list of rhetorical relations is organised into three level hierarchies of 23 fine-grained relations, grouped into seven classes. To automatically learn these relations, we evaluate and reuse features from literature, and contribute three additional features: accusative of purpose, specific connectives and the number of antonym words. We perform experiments on identifying fine-grained and coarse-grained relations. The results show that compared with all the baselines, our model achieves the best performance in most cases, with an accuracy of 91.05%.