Because computational linguistics, or Natural Language Processing, or whatever you want to call it divides historically into two parts: an early period and a late period. The early period–which stretched from the dawn of the computer age to about the mid-1990s–focused on stripping natural language expressions of their particulars and thereby boiling them down to their essence. Beneath the fuzzy accidents of actual English, French, or Russian (and it was Russian back then, it being the Cold War) there lurked the universal core of logical form. In theoretical linguistics you had Montague semantics, which demonstrated for toy problems how to map English to predicate calculus in a way that is faithful to syntactic constituents, programatizing Frege’s idea of compositionality. On the computational side you had SHRDLU written in Lisp which, like, God help us, but at least that’s something you can demo. There’s a lot of impressive work in here that ties in to Reagan-era artificial intelligence research, but ultimately it proves to be something of a dead end.
But thankfully Moore’s Law keeps marching inexorably on, so around the time the Pixies are breaking up naiveté has its day. Natural language research switches from sophisticated logical models to N-grams. Turns out you can make an astonishing amount of DARPA-measurable progress merely by counting word frequency. Did a particular newspaper article come from the sports section or the front page? Just make note of the relative frequency of the words “touchdown” and “Iraq” with the appropriate normalization and you’ll be right ninety percent of the time. It’s amazing. You don’t have to, strictly speaking, understand anything. It’s just a matter of mathematically correlating certain language events with certain other events. And every time you fire a linguist, your performance goes up. No longer are you searching for the linguistic meaning behind the use. It’s kinda like the meaning is the use.
The first of these periods coincides with Wittgenstein’s Tractatus Logico-Philosophicus, and the second with his posthumous “No wait, wait–I had it all wrong” follow-up, the Philosophical Investigations. From what I know of both these works this correspondence is more than coincidence, and more than just a broad alignment of generalities. And I can tell you from personal experience that it’s sure as hell not the case that computer scientists were blinkered by their rabid Wittgenstein fandom. No, I think this is a legitimate case of parallel independent development. And as such us computer scientists might actually have something to learn from Wittgenstein. That, despite the entirely reasonable proposition that philosophy is a gigantic waste of time, in this particular instance someone might have done some of the heavy lifting for us back in the 1920s, and it is worth struggling through a bit of his prose to see if this is the case.