This paper investigates the evaluation of word embeddings trained on a SNCF corpora with a very specific vocabulary. In particular, we consider this task regarding SNCF acronyms and lexicon, and show its hardness. We also highlight the relevance of considering polysemy.