Why computers and AI will probably never be able to fully understand human communications

[Ron Obvious]

I have a friend who is a bit of a Yoga Berra in that he says things in an odd way sometimes, but you usually know exactly what he means.

Yesterday he told me: "I'm really upset with my girlfriend because she deceived me. Last week she told me she'd cheated on me. <pause> Now I find out that that wasn't true".

OK, you know what he means, right? But how would a computer parse that sentence? What's he upset about exactly? Understanding human communications requires up-to-date cultural knowledge, and not just an understanding of grammar and vocabulary because it's so ambiguous.

Compare these two sentences:

Yesterday I saw a movie with my neighbour Fred.
Yesterday I saw a movie with Bruce Willis.

Grammatically, these sentences are the same but they probably mean different things, right? Now, it is just possible that Bruce Willis is a pal of mine I go to the pictures with, and it is also just possible that my neighbour Fred is a movie star. If I were a celeb myself, you might consider either of those statements as likely, but as I'm Joe Ordinary, they're highly unlikely.

Think of how much a computer or robot would have to know to parse those two ordinary sentences correctly. Now what if I'd said:

Yesterday I saw a movie with Marilyn Monroe. This time there's only one way to parse that sentence because Marilyn Monroe is dead. But what if the statement had been made in 1960? Or in 1944 before she was famous? Each such factor that influences the interpretation of the sentence is easy for humans to interpret, but very difficult for computers who don't even know what questions to ask.

For some reason this was something I was thinking about while I lay awake last night.

[Orrex]

1. If an AI can't parse the sentence, then that's our failure, not the AI's

Anyone who's spent time with children has seen the evolution of contextual understanding of language, and if we can construct an AI with a suitable foundation, then that AI will be able to grasp the nuances of language thereafter. It may take a while, but why not? Adult humans frequently mix up context and meaning, so it's only fair that we afford an AI the same leeway.

For that matter, once an AI gets the hang of it, then future AIs could presumably download that understanding straight from the more experienced machine.

If it makes you feel better, you could consider that one AI would be able to chat with another about highly complicated code that humans could never hope to grasp in real time. In that case we would lack both the parsing capability (i.e., "what does all of that mean?&quot and the processing capacity ("there's so damn much of it.&quot

AIs have advanced from nothing to remarkably advanced and interactive in just a few short decades, while language-capable humans have been doing it since at least 1758. Give the AIs a little more time to get it right.

[Ron Obvious]

9. You're more optimistic than I am...

I think you're more optimistic than I am on this. I find it difficult to imagine how all this context and cultural knowledge could be quantified in a way that makes sense to a computer, but it 's quite possible that I'm significantly behind the times here.

[Denninmi]

2. Well, I know this much.

Siri gets a lot of things wrong, and she has an attitude.

[MineralMan]

3. It's going to be very difficult for computers to ever

understand human language. The problem isn't really syntax or grammar. It is a lack of reference points. The computer has no human experiences, and cannot comprehend human references. The sky is blue. I am blue. Even if the computer knows that blue can mean either a color of a mental state, it has no point of reference to use to distinguish between the two statements and understand them. Both are statements that require understanding more than just language. People who are sighted know what color the clear, cloudless sky is during the daytime, at least in most places. The computer only knows those words as data. the second sentence describes an emotional state. Now, I can tell the computer that human beings are not blue in color, but I cannot give a computer a point of reference to understand a human's mental state.

The concepts of language are human, by definition. This becomes clear when trying to use computers to translate between two languages. The use of the concept "blue" to describe a mental state is not a universal usage. In other languages, it may not convey any meaning at all with regard to a human emotional state. Other languages may use a different metaphor to describe the mild depression that being "blue" describes. A human translator can simply substitute a suitable metaphor in the translation automatically, because he or she understands what is meant by "blue" in that usage. A computer is unable to do that. It is unable to differentiate between the two very different meanings in the original language, and will use a variety of matching algorithms to decide what the correct translation will be.

Now, it would be possible to set up rules for the computer that precluded the use of the color meaning when applied to a human, by comparing pronoun lists, but there's a problem. It is actually possible for the color blue to be used to describe a person. Anoxia can result in a blue coloration of the skin, particularly in some tissues. A person can say, my lips are blue. A doctors will know that cyanosis is the medical term for blue people.

That's just a single word that can cause confusion. There are many thousands of such words that have multiple meanings, depending on context. Context requires analysis and comprehension. Computers can analyze, but not comprehend. The problem will remain.

[Odin2005]

4. I think this boils down to hardware differences between brains and computers.

Brains work based on context and association and are massively parallel in their processing, while electronic computers work based on linear if-then logic gates.

[MineralMan]

5. Exactly. Language is pretty much an analog thing,

rather than a digital one. Everything about language requires a complex, inexact process, to understand.

Humans misunderstand each other pretty well, too, so language is not something precise that is easy to process. I'm not sure it's a parallel or binary process, but I know that computers make a pretty good hash out of human language.

Now, I love language, and I love computers. That's why I loved programming. It's just another language, with rules and restrictions. I taught myself both QuickBasic and Visual Basic. I know those two languages aren't much respected, but I actually created a number of pretty popular consumer applications with them. What I liked about both was that they seemed more language-like to me. I'm a weird sort of coder, but Basic suited my style very nicely. When Microsoft finally changed VB to be closer to a Pascal type of language, the fun went out of it for me. Too much structure and too little connection to human language for me. I missed my global variable arrays too much.

So, I basically shut down my little shareware software company, dumped the registered versions out as freeware, and moved on to other pursuits. By now, few people are using any of my stuff, but I still answer the odd support call from one of my old registered users and a few of the programs are still circulating out there and being used.

[Evoman]

14. What sort of programs did you make?

Just curious. What was the name of your software company?

Do you still do any programming?

[MineralMan]

15. It was a variety of stuff.

The company was OsoSoft Shareware. I did Rockford, a business card design program (the first one) for creating and printing sheets of business card. Another was MultiLabel, a label design program that supported all Avery formats, and had some nifty features like a contact database and automatic font sizing that would fit too-long text on the label by reducing font size 0.1 at a time until it fit. There was Fonter, a Windows 3.1/95 font utility that included a font installer and previewer, along with the ability to print font cards that displayed all the font's characters in a nice format. Prompter, a program to automatically or manually scroll text for speeches, TV news rooms, and other stuff. That one included two fonts I created, one which would display pronunciation marks and other that displayed text in mirror format. The program would automatically reverse its display, for those who used mirrors on the teleprompter. The fonts are still circulating. I did another version of Prompter, designed especially for low vision people or people with other limitations. It served as an automated reader application, with very simple controls that changed scroll rate, etc. That one I did for the public domain. Both could import text from almost any source, and could accept any word processor documents or just about anything else with text in it.

There were lots of others, too. Some were shareware, but I put out a lot of freeware, too. I also found ways to extend the capabilities of both QuickBasic and the early versions of Visual Basic, and made all those things available freely. Pre-Windows, I had a bunch of apps out there, too.

I don't do any programming any longer. I've moved on to other stuff. A search for OsoSoft shareware brings up a lot of old, old links.

[Ron Obvious]

11. Absolutely...

Programming and linguistics are two of my fascinations as well, though I've not met too many people who shared both these interests before.

Human language conveys so much more meaning than the words themselves which are rooted in history, culture and value judgements. I think you'd need to be fluent in more than one language to really understand this. I found that most monolingual people don't really understand that translation is a subtle art and not just the mere substitution of words and reordering of grammar.

There are the cultural prejudices that attach to words which are usually not mentioned in the dictionary. Calling a girl 'pretty' is a compliment, but calling a boy 'pretty' may or may not be depending on the people involved and their value judgement and may imply a lack of masculinity. Likewise, a 'handsome' woman sounds severe rather than attractive, at least to my ears. Handsome also implies the first bloom of youth is gone, whereas 'pretty' implies youth. Dictionaries don't mention this usually, though.

Additionally, English is dangerous to interpret in particular because so many words do double-duty as nouns and verbs.

There's the classic: "Time flies like an arrow but fruit flies like a banana", of course, but books like Anguished English abound with examples of English ambiguity. I've always liked this one:

"American Ships Head to Libya".

Of course we know it's about warships on their way to a troublesome part of the world, but how is a computer supposed to understand that this isn't about a disturbed individual mailing a package containing a human head?

I suppose we would have to assign probability values to every possible interpretation, but language changes all the time and this would have to be kept up-to-date.

As you said, the problem will remain.

[pink-o]

12. And the cultural differences go deeper than a good translation.

I speak Spanish, and I also live in San Francisco in a very small house. So I've said to my Latino friends: "Mi Caja es Su Caja" (Caja is a box) which would be an ironic, self-depricating pun in English. But Spanish speakers don't have the same double-meaning word sensibilities as English speakers. We use the same word for so many parts of speech and so many sound-alikes. I haven't encountered another language where that is the case, nor another culture that uses humor about it the way English speakers do.

[MineralMan]

13. Yes. Cultural sensitivity is crucial

for doing translations. Only a native-speaker of the target language who is fluent in the source language can properly take those into account in a translation. Computers will never have that ability.

[Tuesday Afternoon]

6. I like this thread.

I like this post.

wait. what.

[Grantuspeace]

7. It can be dangerous to dangle your modifiers!

The classic, "I saw an elephant in my pajamas" comes to mind.

[Ikonoklast]

8. Small elephant.

Or some big-ass pajamas.

[JustABozoOnThisBus]

10. The pajamas are so big ...

... they need to be packed in a trunk