Enhancing pre-trained contextual embeddings with triplet loss as an effective fine-tuning method for extracting clinical features from electronic health record derived mental health clinical notes.