Most commenters here find these demos very underwhelming and I must agree with them. Well known open source NLP packages like Stanford NLP or NLTK do much better on these (very academic) tasks.

I think the problem is that Watson's demos have been trained for very specific and narrow domains (like "cloud services" in the case of Resonance). What we really want to know is how easy (or difficult) it is to train Watson for our own domain.

If Watson has to make assumptions on my domain, then it will be really hard to build something useful with it.