Training Hybrid Language Models by Marginalizing over Segmentation - View it on GitHub