I'm very skeptical that this would work outside of toy examples, though it depends on what is meant by language-independent AST. For example, the best way to t

I'm very skeptical that this would work outside of toy examples, though it depends on what is meant by language-independent AST. For example, the best way to translate Spanish "X dió un golpe a Y" would be "X hit Y". But my naive idea of what the AST for the Spanish sentence would look like would be something like `(GIVE (X HIT Y)`, which when naively transduced to English would be the "X gave a hit to Y", which is either unidiomatic or means the wrong thing altogether. In order to avoid this problem, the AST would have to be a more abstract representation of the semantics. And coming up with a sufficiently expressive, tractable, and neutral representation of natural language semantics is an unsolved problem that people are still devoting their whole careers to. I was briefly involved in a very early stage startup that was considering using systems like this for better machine translation. We ran into problems like the above, and also: ambiguity, and the fact that the hand-written grammars and semantic representation systems were just very brittle and incomplete.