—The difficulty to deal with new users, items and the poor explainability of predictions are well-known weaknesses of collaborative filtering. Classically, the cold-start issue is tackled either by asking for user interaction or exploiting side information while additional explanations are often extracted afterwards in a standalone process. Here, we propose a text-based collabora-tive filtering recommender system which provides a framework to solve both issues. Our method extends a scalable text embedding technique to build a unified vector space where users and items are mapped. We show how to use this space in a collaborative filtering scheme and demonstrate the interest of our text-based method for cold start soothing and recommendation explanation. The suitability of our approach is backed by competitive results on rating prediction task.