Solving natural language query is at least as hard as asking the system to write a program for an arbitrary goal. So if you can ever build a program that can take an arbitrary goal and output the program source code then you can also solve arbitrary natural language queries efficiently.
So most "supposedly" AI systems do brute force or probabilistic templating - i.e. mapping the query to a known structure of sentences. Then you get translation of these known structures of sentences to known machine level structured queries. This is why systems like Siri would fail on its face as soon as you ask even mildly "hard" question. For example, queries that starts with "How many..." such as "How many teeth humans have?" are very easy to solve and in fact it can be your weekend project for Wikipidea corpus with impressive accuracy. But if you change this query to "What is the factorial of the number of teeth that human have after subtracting number of legs octopus has?" then you would get no where.
So most "supposedly" AI systems do brute force or probabilistic templating - i.e. mapping the query to a known structure of sentences. Then you get translation of these known structures of sentences to known machine level structured queries. This is why systems like Siri would fail on its face as soon as you ask even mildly "hard" question. For example, queries that starts with "How many..." such as "How many teeth humans have?" are very easy to solve and in fact it can be your weekend project for Wikipidea corpus with impressive accuracy. But if you change this query to "What is the factorial of the number of teeth that human have after subtracting number of legs octopus has?" then you would get no where.