This right here is the right strategy, make a new framework and wait for the model to be re-trained with a corpus large enough for it to be correct. Rinse and repeat.