Pre-Trained Embeddings for Enhancing Multi-Hop Reasoning