Minor Description
The class BratDocument's static parseDocument generated SpanAnnotations with empty covered text… I think that is a little messy. was this done in order to save memory? Otherwise I think you should rewrite the getCoveredText so it looks like opennlp.util.Span's CharSequence Span.getCoveredText(CharSequence text)
included a patch that compiles on current trunk.
We could also fix SpanAnnotationParser to read the "covered" text but I think I should look more like the opennlp api.