#1 is that people don't want to do the effort to make training sets specific to their purpose.
The table stakes for a text analysis project are to annotate from 3,000 to 30,000 examples. A well caffeinated, motivated and trained annotator can annotate maybe 2,000 examples a day so this is very feasible but people would rather spend the person*days screwing around with Tensorflow, NVIDIA drivers, BERT, and whatever is the new shiny rather than do the work which is on the critical path.