It sounds to me as if you have done the first (easy) half of your choice already: You want to do something that helps in the public sector and makes use of AI / natural language; and has the potential to be beneficial to many people. "But I don't know what they need". What do you do when you have a question about what some group needs? Research. Have you done that research? What do they (think they) need?
Just for the fun of it:Something that reads traffic and store signs, translates them, and speaks them in any language.
Something that helps poor people cope with government forms
A tool that takes phone calls, emails, printed material and consolidates it all into a report about which parts of the city need what services, based on complaints.
A much much better phone answering system instead of one of those horrible phone trees: Let me explain what I want and the tool directs me to the right place. This would be a bad idea because it takes away a job, except that job is already gone, and being done so poorly already.
A tool that tells technicians what the problem is likely to be, and points them to the likely location of the problem. (Phone or networking outages, for instance). Speaking so they can use their eyes on the job while listening
A robot tutor for one or more subjects
A device that generates "good" test questions for exams, possibly based on how well answers to past questions matched the quality of the students who answered them. (SAT already does this, but not automatically)
By the way: Do you see how I was creating these ideas? I started with "Hmm, what makes use of natural language processing that I've never seen?" and then each idea leads almost immediately to the next.
