To BERT or not to BERT - Comparing Contextual Embeddings in a Deep Learning Architecture for the Automatic Recognition of four Types of Speech, Thought and Writing Representation.