Improving astroBERT using Semantic Textual Similarity