Exploring the Value of Pre-trained Language Models for Clinical Named Entity Recognition