Collapse of Self-Trained Language Models