From Zhang et al. (2015): "Discourse relation recognition (DRR) aims to automatically identify the internal structure and logical relation- ship of coherent text. [...] implicit DRR remains a serious challenge because of the absence of discourse connectives."
An example from Penn Discourse Treebank:
Arg1: (But) our competitions say we overbid them Arg2: who cares
The connective "but" was manually inserted to indicate comparison relation.
- Zhang et al. (2015) uses two convolutional neural networks over two candidate text, followed by one hidden layer and a softmax layer.