Create some documentation in README for the different special tokens that we recognize (, , ,etc., where "recognize" just means they aren't wordpieced), and for what types of data they are used for. This should also clarify that the --token argument currently only applies to for entity classification and not , for relations.