I tried this, and while 'targeted sentiment' may be neutral, 'document emotion' has Joy at 0.880435 vs the rest at 0.

If you change your example from 'I like it' to 'I like puppies', the 'targeted sentiment' changes from neutral to 'puppies positive 0.550431'.

While the first example you found does seem odd, the rest seems to work quite well. Maybe it just has problems attaching 'it' to a 'target'.

It has trouble with negation outside the relevant clause. So it is not fooled by e.g. "I don't hate it" vs. "I don't like it", but the first of the following two sentences is rated more negative than the second:

I don't seem to hate it.

I don't seem to like it.

Edit: In fact it can even be fooled just by increasing the linear distance between a positive term and negation. The following is rated +0.6!

I don't in any way shape or form like it.

I think this is actually very accurate as to how humans use the sentence "I like it". If you cook dinner for someone and you ask how it is and they say "I like it" it generally means it's not bad but also not good.

Edit: My suspicion was confirmed when I tried "I like it a lot" had a positive sentiment with 0.711.

"I couldn't live without it" is rated -0.6.

Google has a similar API:

https://cloud.google.com/natural-language/

Microsoft too: https://azure.microsoft.com/en-us/services/cognitive-service...

Alchemy was an acquisition.

It was an odd acquisition considering IBM Watson's supposed sophistication. Why not just use Watson's components rather than acquiring a company that did roughly the same thing?

Which components? The Jeopardy-Watson didn't have any components immediately relevant to targeted sentiment analysis; it would make sense to acquire the missing components externally.

Quite a few seemingly similar AI/natural language/etc problems are actually very separate.