1
$\begingroup$

We know a lot of common sense about the world. Things like "to buy something you need money".

I wonder how much of this common sense comes about through actual someone explicitly telling you the instructions "You need money to buy things". Which we store in our brains as a sort of rule. As opposed to just intutively understanding things and picking it up.

I am imagining children playing at shop-keeping and saying things like "I give you this and you give me that". And other children not quite understanding the concept of buying things until being told by a teacher.

If so, giving a computer a list of common sense rules likes these is no different to teaching a child. So I am wondering why this area of AI research (semantic webs etc.) has been frowned upon in the last decade in favour of trying to learn everything through experience like deep neural networks?

$\endgroup$
  • 1
    $\begingroup$ There are way too many rules and too many exceptions to the rules and very much that is context-dependent in a nuanced way. $\endgroup$ – George White Jan 13 at 16:27
  • 1
    $\begingroup$ @GeorgeWhite is right. Also the ability of deep models to produce relatively higher accuracy has driven researchers away from conventional AI. With growing amounts of data, its easier for models to identify patterns and "predict" words instead of actually "understanding" and "writing" words. $\endgroup$ – Sharan Jan 28 at 5:39
  • $\begingroup$ Yes, there are a lot of rules. But that doesn't stop teachers from trying to teach things like science. They don't just give up saying "There's too many facts!" $\endgroup$ – zooby Jan 29 at 2:13
1
$\begingroup$

If I understand correctly, what you are looking for is called "common sense reasoning" in NLP research.

Research in this field revolves around benchmark data sets, where good performance indicates some ability to do common sense reasoning. Here is a nice collection of data sets and research by Sebastian Ruder:

http://nlpprogress.com/english/common_sense.html


In the end, the main question is not

How much knowledge of the world is learnt through words?

since it is virtually unanswerable if asked in this form. A question that is answered in NLP common sense reasoning research is

Out of 100 specific decisions that my model needs to take, for how many does the model show the ability to reason correctly?

| improve this answer | |
$\endgroup$
  • 1
    $\begingroup$ I wonder how much of this is statistically based. A human can reason like this "All men are motal, Socrates is a man, therefore Socrates is mortal". This makes perfect sense to a human. But to an AI it seems like it can only reason like "therefore Scorates is mortal with 95% probability". What seems to be required is a way to manipulate words in the brain much like making moves in a chess game. The same mechanism that an AI can use to think about future moves in a chess game should be used in reasoning. $\endgroup$ – zooby Jan 29 at 2:17
0
$\begingroup$

Grounded language learning

Your description does not match the current understanding of how children learn language, and giving a computer an explicit textual list of common sense rules is very different from teaching a young child. (Some aspects of teaching older children is more similar to that, but by that time they fully know the language already).

Much of language is acquired through interaction (both sensory and motoric) with the physical world, applying words together with a shared focus of attention to something real. I.e. you talk about a cat or a ball and it's behavior while both you and the child are paying attention to that behavior or object or an image of it. The same applies later for more complex topics such as social situations - to teach a topic, to a child, you'd inevitably use a shared attention to specific events in the real world or specific events in a mostly shared 'model world' i.e. one reconstructed from memory (what you saw your sister do five minutes ago) or imagined/hypothesised (if you do this, these consequences might happen).

Attempts to replicate this process in artificial systems is usually called 'grounded language learning', and there's extensive published literature on that which may be interesting to you.

In essence, the assumption is that English (or any other) words to an artificial system are just as useful as a Chinese-Chinese explanatory dictionary to me. If I speak some basic Chinese, then I can use that dictionary to expand my vocabulary and understand complex Chinese - but if I have no grounding in Chinese whatsoever, then the expectation is that it's impossible to reconstruct a language from that data alone.

| improve this answer | |
$\endgroup$
  • $\begingroup$ That's all very well if you want to create an AI like a young child not an adult. But also children go to school and are taught facts and algorithms (like long division) by teachers. So one needs both. Lots of behaviour is learned by words "don't do that", "that's a zerbra", "wash your hands before dinner", "hold your fork in your left hand", "don't chew with your mouth open" $\endgroup$ – zooby Jan 29 at 2:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.