Taming Pre-trained Language Models with N-gram Representations for Low-Resource Domain Adaptation