IMO we are still lacking a convention to represent illegible substrings. DTABf (TEI) uses gap for this.
Since there is a dependency from GT to OCR training to OCR inference to OCR postcorrection, we should make this as concrete as possible without breaking existing habits. For example, in public GT datasets you often see € or £ to represent this directly in the string. The downside is obviously that you might somehow end up confusing these substitutes with their actual meaning after all.
If possible we could also try to enforce a non-printable character like ASCII bell, substitute or unit separator. In the simplest form, we just use the empty string – but that only works when transcribing on character level, and OCR is trained on line level.
IMO we are still lacking a convention to represent illegible substrings. DTABf (TEI) uses gap for this.
Since there is a dependency from GT to OCR training to OCR inference to OCR postcorrection, we should make this as concrete as possible without breaking existing habits. For example, in public GT datasets you often see
€
or£
to represent this directly in the string. The downside is obviously that you might somehow end up confusing these substitutes with their actual meaning after all.If possible we could also try to enforce a non-printable character like ASCII bell, substitute or unit separator. In the simplest form, we just use the empty string – but that only works when transcribing on character level, and OCR is trained on line level.