In this preliminary study, we experiment with the use of DBnary, a big lexical knowledge graph, to create word embeddings that could be used in NLP downstream tasks. Our gamble is that word embeddings created from lexical data (instead of language corpora) may exhibit less biases while still being usable as the first layer of deep learning approaches to NLP tasks. We tried very basic method of embedding creation from lexical graph and evaluate (1) the intrinsic performance of the created embeddings on word similarity and word analogy test sets and their extrinsic quality in POS tagging and NER downstream tasks, along with (2) the biases they may exhibit. Such embeddings show promising performances outperforming word2vec on few specific tasks, while still not on par on most others, but we confirm that they exhibit less bias overall.