Listening to Groucho Marx make jokes can teach us a lot about AI

We hear a lot about AI these days and now more and more about NLP (natural langue processing.) Apparently computers can understand what we say. Uh wait. Not so fast. 

I learned, after 40 years of working in NLP, how hard it is to understand natural language. The best way I have to explain this to people who think chat bots will soon replace people is through jokes. I especially like Groucho Marx jokes, so here are a few of his favorite lines:

Outside of a dog, a book is man's best friend. Inside of a dog, it's too dark to read.

While hunting in Africa, I shot an elephant in my pyjamas. How an elephant got into my pyjamas I'll never know.

We took pictures of the native girls, but they weren't developed. . . But we're going back next week.

Can a computer understand these jokes? Not today, nor any time soon. Why not? Partly because many of Groucho’s jokes depended upon ambiguity. Words mean different things in different contexts. People, when listening to others speak, make assumptions and predictions about what will come next, based on what they know about the speaker, and the world, and act accordingly.

Here is a sentence to think about:

The old man’s glasses were filled with sherry.

As we begin to hear or read this sentence we create an image in our minds. We don't always realize that we are doing this, but we image an old man with glasses, and then we have to adjust our image. This kind of thing is exploited by comics like Groucho, but it happens every time we hear anyone say anything. 

A person determines for “the picture of native girls” joke, that the “they” in “they weren't developed” refers to film and then realizes that it refers to “the native girls.” We realize that because we know something that is not in that sentence in any way, namely that “native girls” often were not wearing shirts in pictures in National Geographic at the time of the joke. You can’t laugh without thinking about all of that. Similarly “in my pyjamas” causes us to create an image in our minds that the next line makes us realize was the wrong image.

Similarly, “outside of a dog” forces us to disambiguate “outside” to mean “apart from” which we then need to reparse when we hear the phrase “inside of a dog.” We laugh when we have to re-image. 

Current NLP programs would not only not laugh at these, they wouldn’t get confused, nor would they have to re-parse, when they realized they had made a mistake because they would never realize they had made a mistake.

Watson, that bastion of NLP, did the following the other day during an NBA game where it was “helping” the broadcasters:

“My trade off analytics indicate that no one creates more space on offense. This allows him to nail a jumper from a densely populated urban area.”  

https://www.inverse.com/article/33026-ibm-watson-basketball-commercial

Apparently Watson was trying to say “downtown.” It seems safe to say that Watson knows nothing about language at all. What its knows about is words, especially how to look them up and count them.

Would Watson understand Gorucho when he said:

I have had a perfectly wonderful evening, but this wasn't it.

or this: 

No man goes before his time - unless the boss leaves early.

Why not? Because the first sentence requires a knowledge of social etiquette and things people say to each other to be nice. The second requires understanding that the phrase “ going before your time” means dying young. How would Watson know that? Someone would have to have had to explicitly program that in. There are many expressions like that and that would mean a lot of programming, but let’s assume Watson would be able to match that phrase to its data base of phrases. It still wouldn't get the joke because it would not know that going before your time could refer leaving the office. It wouldn’t know that because no one would have thought to program that into Watson because if it tis a rather weird thing to say, nevertheless we can interpret it and then laugh at the joke. People can figure stuff out. So-called “AI’s” cannot.

My last Groucho remark is one of his most famous:

I refuse to join any club that would have me as a member.

What does this even mean?

It is a commentary on exclusionary tactics of private clubs in Groucho’s era, less prevalent now, but quite common in Groucho’s time. And it is a commentary on Groucho’s self-assessment of his own character, which is why it is funny.

On that same topic when his 8 year old daughter was barred from a club where her friend had brought her to swim because they didn't allow Jews, he said: 

"She's only half Jewish. How about if she only goes in up to her waist?"

Watson won’t be getting that one anytime soon. And that is the real issue. Understanding natural language involves having a detailed model of how the world works, why people do what they do, and how one makes plans to counteract things that others do that you don’t like. Groucho was kidding but others might have had a different and angrier response. Watson wouldn't get angry nor would it try to respond, because it is not a thinking entity with a world view and a knowledge of how the world works.  Counting words won't get any computer to have any idea what Groucho’s remark was even about.

Kevin Kelly of Wired Magazine said the following:

The business plans of the next 10,000 startups are easy to forecast: take X and add AI” — Kevin Kelly, Wired

He left out “say AI and not understand what the heck you are talking about," which is pretty much what venture capitalists and those taking their money are doing today. Here is what a VC (John Henderson (Whitestar Capital) had to say:

For the most part, recruiters are essentially intermediaries that connect candidates and employers based on a defined set of features (location, industry, skills, experience). I can’t wait to meet a company using machine intelligence and NLP to identify and curatecandidates automatically. The first interview will almost certainly also be conducted by a computer.

Great idea. Let's not try to understand what a person is all about. Instead, let’s just count how many times he says AI in his resume.