Towards Concept-Aware Language Models