Application of a Domain-specific BERT for Detection of Speech Recognition Errors in Radiology Reports