STARLING: Self-supervised Training of Text-based Reinforcement Learning Agent with Large Language Models