In this work we propose a pretraining procedure that aligns a graph encoder and a text encoder to learn a common multi-modal graph-text embedding space. The alignment is obtained by training a model to predict the correct associations between Knowledge Graph nodes and their corresponding descriptions. We test the procedure with two popular Knowledge Bases: Wikidata (formerly Freebase) and YAGO. Our results indicate that such a pretraining method allows for link prediction without the need for additional fine-tuning. Furthermore, we demonstrate that a graph encoder pretrained on the description matching task allows for improved link prediction performance after fine-tuning, without the need for providing node descriptions as additional inputs. We make available the code used in the experiments on GitHub( https://github.com/BrunoLiegiBastonLiegi/CLEP ) under the MIT license to encourage further work.