Discourse connectives play an important role in making a text coherent and helping humans to infer relations between spans of text. Using the Penn Discourse Treebank, we investigate what information relevant to inferring discourse relations is conveyed by discourse connectives, and whether the specificity of discourse relations reflects general cognitive biases for establishing coherence. We also propose an approach to measure the effect of a discourse marker on sense identification according to the different levels of a relation sense hierarchy. This will open a way to the computational modeling of discourse processing. 1