ContentVec is being used incorrectly

**Describe the bug**
https://github.com/auspicious3000/contentvec/issues/6

**Additional context**
Models trained with the correct annotation that fixes this part will be incompatible, so tweaking config. There is a hypothesis that with luck, fine tuning will solve the problem.
