Certainly it's up to the vagaries of fate and such, but when I was in school I worked in a lab that for several years got fairly reliable DARPA funding for NLP work, and while they never built the department around it or anything it was nice to have that consistent year over year assumption that one or two projects would be interesting to the DARPA folks.
For them (us?) it was more a matter of "well this doesn't necessarily merit or reach the requirements for quote-unquote 'real' research, but it's really interesting and you know some investigation couldn't hurt". Those got pitched to DARPA, they asked "can you do it with Arabic and Chinese too?" and we said "no" and they funded us anyway a few times out of ten. Not bad, really.
For them (us?) it was more a matter of "well this doesn't necessarily merit or reach the requirements for quote-unquote 'real' research, but it's really interesting and you know some investigation couldn't hurt". Those got pitched to DARPA, they asked "can you do it with Arabic and Chinese too?" and we said "no" and they funded us anyway a few times out of ten. Not bad, really.