This work addresses the use of computational linguistic anal- ysis techniques for conceptual graphs learning from unstruc- tured texts. A technique including both content mining and interpretation, as well as clustering and data cleaning, is introduced. Our proposal exploits sentence structure in or- der to generate concept hypothese, rank them according to plausibility and select the most credible ones. It enables the knowledge acquisition task to be performed without su- pervision, minimizing the possibility of failing to retrieve information contained in the document, in order to extract non-taxonomic relations.