Extracting formal knowledge (ontologies) from natural language is a challenge that can benefit from a (semi-) formal linguistic representation of texts, at the semantic level. We propose to achieve such a representation by implementing the Universal Networking Language (UNL) specifications on top of RDF. Thus, the meaning of a statement in any language will be soundly expressed as a RDF-UNL graph that constitutes a middle ground between natural language and formal knowledge. In particular, we show that RDF-UNL graphs can support content extraction using generic SHACL rules and that reasoning on the extracted facts allows detecting incoherence in the original texts. This approach is experimented in the UNseL project that aims at extracting ontological representations from system requirements/specifications in order to check that they are consistent, complete and unambiguous. Our RDF-UNL implementation and all code for the working examples of this paper are publicly available under the CeCILL-B license at https://gitlab.tetras-libre.fr/unl/rdf-unl