Community Question Answering (cQA) continues to gain momentum owing to the unceasing rise of user-generated content that dominates the web. CQA are platforms that enable people with different backgrounds to share knowledge by freely asking and answering each other. In this paper, we focus on question retrieval which is deemed to be a key task in cQA. It aims at finding similar archived questions given a new query, assuming that the answers to the similar questions should also answer the new one. This is known to be a challenging task due to the ver-boseness in natural language and the word mismatch between the questions. Most traditional methods measure the similarity between questions based on the bag-of-words (BOWs) representation capturing no semantics between words. In this paper , we rely on word representation to capture the words semantic information in language vector space. Questions are then ranked using cosine similarity based on the vector-based word representation for each question. Experiments conducted on large-scale cQA data show that our method gives promising results.