Using VADER Sentiment Analysis makes distributions overlap : how to improve my model

I use VADER Sentiment Analysis on a customer reviews dataset.

VADER breaks down feelings of satisfaction and dissatisfaction into neutral and positive negative components. Plotting the distributions, I see that those of satisfied and dissatisfied customers overlap quite a bit.

I would like to know if I can improve my model: I imagined training it only on non-overlapping dataset values, is this a correct method? Is there another method?

I am a beginner, please be cool, thanks again for your help.