Me [a computer] Talk Pretty One Day

Me [a computer] Talk Pretty One Day

Why language so easy for us and so difficult for computers

Consider the following sentences (example taken from Gibson and Warren 2004):

• Who did the consultant claim that the proposal had pleased?
• Who did the consultant wonder which proposal had pleased?

If I asked you to say which sentence was well-formed, you’d probably say the first sentence is and the second isn’t. In linguistics, we would say that the first one is grammatical and the second one isn’t. However, grammaticality is not always as simple as “this one works” and “this one doesn’t work.” In many cases, including my own BA thesis, people will be asked to rank the grammaticality of a sentence on a scale (in my research, I asked subjects to rate sentences on a scale from 1 to 7, which is relatively standard).

The reason I bring up this concept of grammaticality is to highlight an important aspect of language, its ambiguity. People could look at the same exact sentence and come up with very different judgements. One might say “yeah, that sounds like a perfectly fine sentence,” another might say, “that sentence makes absolutely no sense,” and another might say, “well, in a certain context, or if there’s a certain stress pattern, it could make sense.” While we certainly have some rules concerning what does and does not create a well-formed sentence, we can’t always say for certain if a given sentence unambiguously violates or does not violate a rule. Additionally, there are circumstances where a sentence may not violate a rule, but the structure is such that it has a higher processing cost. This also why linguists are still trying to determine what the rules are for sentence production.

Language is a lot more complicated than we give it credit for. Not only is there ambiguity in the way we interpret words and sentences (i.e. lexical and structural ambiguity), but also ambiguity in what constitutes a rule violation. That said, why does it not seem that complicated to us on the surface? These constraints show why it would be difficult for a computer, a machine that deals in absolutes and numbers. So, why does it come so naturally to us?

You may or may not have seen this tweet at some point if you, like me, spend way too much time looking at memes. Now imagine, you’ve got an Olympic figure skater, doing leaps and turns and being generally graceful for their entire dance. Then you have someone like me, who hasn’t ice skated in years and was laughably terrible in her ballet classes, clinging to the walls on the perimeter and probably faceplanting the moment I try to do some kind of jump.

Of course, the reason the figure skater could easily do a toe loop and I barely even know what that would look like is because the figure skater has been training for most of their life. In other words, a complicated maneuver is simple for them because they’ve been doing that for so long. This is Moravec’s paradox , the stuff that is the simplest for us is the most difficult for computers. Language production and comprehension is simple for us because we’ve been talking for thousands of years. On the other hand, we’re just starting to teach computers to do just that.

In one of my previous posts, I talk about how we should use AI to assist us, not replace us . I discuss a few reasons why, but another big reason is that we really don’t give ourselves a lot of credit for what we humans are capable of. Just because things like language come easily to us doesn’t mean it is easy, a lesson we’re learning in the field of NLP. Our brains have been molded and have evolved over thousands of years. While, of course, we want to innovate and advance technology, we should not aspire to completely replace humans because that would require a massive amount of work. Why reinvent the wheel when we have a perfectly good one here? If we want to efficiently and productively spend our time and resources, let’s build things to work with the wheel, not replace the wheel.

[1] Gibson, Edward, and Tessa Warren. “Reading‐Time Evidence for Intermediate Linguistic Structure in Long‐Distance Dependencies.” Syntax 7, no. 1 (2004): 55–78.

[2] Sennet, Adam, “Ambiguity”, The Stanford Encyclopedia of Philosophy (Spring 2016 Edition), Edward N. Zalta (ed.), URL = <https://plato.stanford.edu/archives/spr2016/entries/ambiguity/>.

[3] Hamer, Ashley. 2018. “Moravec’s Paradox Is Why the Easy Stuff Is Hardest for Artificial Intelligence.” Curiosity.com. June 11, 2018. https://curiosity.com/topics/moravecs-paradox-is-why-the-easy-stuff-is-hardest-for-artificial-intelligence-curiosity/.