Our research work focuses on the extraction of semantic relations essential to the analysis of discourse, in which links are made between units of text in virtue of semantic information. These relations are frequently implicit and require for their identification inference using lexical and compositional semantic information. Our approach addresses this challenge, using non annotated data with automatically detected features to find adjacent clauses in which various discourse relations occur, to allow the extraction of typical lexical features. Specifically, when these clauses contain certain discourse connectives, we recover automatically a discourse relation that we associate with the main verbs of those clauses. We extract triples consisting of the two verbs and a semantic relation from a large corpus with the aim of inferring that such a pair of verbs can suggest the semantic relation even in the absence of an explicit discourse marker. Altough this analysis allows for noisy results, its application on a large corpus yields significant triples. Thus we compute significance measures to build a resource composed of ranked triples. This resource is evaluated first intrinsically by computing the correlation between our significance measures and a human association value obtained by manual annotation for selected triples, then extrinsically by computing coverage on discourse annotated corpora, and finally by including our triple significance scores as additional features in an automatic discourse relation prediction model. Our positive results show the potential impact of our resource for discourse analysis tasks as well as other semantically oriented tasks.