Improve Microsoft cognitive services sentiment analysis?

Question

I find sentiment analysis to be frequently off, completely off, of what the sentiment actually should be. Here are some examples,

1. I told you to not do that - Sentiment 0.81 (positive) .. how is this positive?
2. if you drop the glass i will beat you - Sentiment 0.74 (positive) .. how in the world is beating someone positive?
3. You made a mess - Sentiment 0.76 (positive) .. how is this positive?

etc. many other such examples!

Questions,

• Is it possible to tweak this per your model?
• Is it possible to have a conversation thread be analyzed for sentiment? For instance, utterance: you made a mess, it's okay, count to ten .. .. here count to ten is .. reassuring .. i.e. it's okay, take a deep breath. Sentiment should be positive!

But saying just "count to ten" should be not as positive.

How far can I take Microsoft cognitive services sentiment analysis? Are there better/other options?

Thanks!