Quote from rufus_4000:
I know a system run by a huge treasury trader that downloads the Bloomberg headlines using the Bloomberg API (client side), so that it would react to economic events a fraction of second faster than human traders. The system just do some very simple text parsing on the economic announcement (since they are basically formulaic), and determine what the direction of the trade. So it avoids any kinds of semantic understanding, AI, linguistics, etc.
It was an interesting setup, but the system is watched by his clerk all the time, in the event that the system would not interpret the announcement correctly. But the trader was huge enough that a +/- $50k on the event don't make that big of a different to him. And if he is right >= 60-70% of the time, he is happy.