Here is a personal anecdote that will hopefully make clear what I'm talking about. When I first used LDA I was blown away at how context could disarm homonyms.

Here is a personal anecdote that will hopefully make clear what I'm talking about. When I first used LDA I was blown away at how context could disarm homonyms. This type of insight isn't exactly mind blowing after the fact, but it's still compelling enough to expect that statistical models produce insights that are broadly applicable without having to resort to proving it.