The Prospects for AI

John Haugeland: 'Understanding natural language' Lycan (21).

Some comfort for the sceptics this week, although they don't need it I know.

I want to draw before you some of the reasons why we seem a long way off from having established that the human brain is a computer, or, to put it another way, I want to rehearse some of the conditions that will have to be met if that claim is ever to be maintained.

They are spelled out by Haugeland.

Natural Language capacity a good test for intelligence

They are cast in the form of outstanding difficulties in the implementation on a computer of natural language - getting a machine to speak English, say.

The idea of getting a machine to speak English has always seemed a good one, and has almost always been grossly underestimated, even by the brightest and best. Alan Turing wrote in his 1950 paper that in order to teach a computer English, 'things would be pointed out and named, etc.'. This has been called 'one of the most breathtaking 'etceteras' to issue from a responsible pen.' (Me, TM, p.243)

Haugeland argues that the points he is going to make, about the immense difficulty of getting machines to 'talk proper', are not just about language use, but offer a good test of the general 'intelligence' of a machine, a good test of success for AI in general.

One reason is that understanding a text requires understanding its topic.

[Another reason why it is a good test is the clarity with which it can be applied. It poses the question: does the text enable the candidate to answer those questions it would enable competent people to answer?

Haugeland, in Lycan, p.660.]

Four varieties of holism

Haugeland speaks of 'holism' as a way of insisting that understanding what someone is saying cannot be broken down into the sum of understanding component bits of what they are saying.

This runs counter to the naive idea that understanding as sentence is a matter of the sum of the understandings of each of the words.

Think of an example.

'It would be nice to have a mince pie after the seminars on Thursdays.'

On a very simple model of how language works, you would think this could be translated into Spanish say by taking each word in turn, looking it up in the dictionary and writing down the Spanish word.

Everbody knows it isn't that simple. One thing is this. You often need to take the whole sentence into account before you can see what it means.

Just one difficulty is this. Lots of words are ambiguous. You can only work out what the relevant meaning is when you know the context.

When we encounter 'mince' for example, are we to understand fragmented meat?

So in a sense you can't translate piecemeal. It has a 'holistic' dimension.

1. HOLISM OF INTENTIONAL INTERPRETATION

When you interpret someone as acting on a belief or desire you have to treat them as a single system possessing a large interlocking set of beliefs and desires.

Attributing someone a particular belief is done on the basis of assuming they have all kinds of other beliefs. You decide what is reasonable for them to have as a belief in their circumstances in the light of what other beliefs you assume they have.

Somebody lies down in the middle of the road.

What belief / desire do we attribute to them?

Because they feel uncommonly tired?

Because they think they are a strip of asphalt?

Because they are protesting against heavy lorries?

Which of these beliefs we attribute to an actual person going in for this type of behaviour will depend on what other beliefs we attribute to them.

Eg if we attribute to them the belief that they think they are in their bedroom, we may think they are just tired.

Intentional explanations of behaviour - explanations invoking beliefs and desires - are holistic.

Haugeland wants to pick this sort of holism out only to distinguish it from:

COMMON-SENSE HOLISM

You need to apply 'common-sense' in order to understand a sentence: but which bits of common-sense is horribly unpredictable.

Take as an example:

'The box was in the pen'

(playpen, cattle pen)

'I left my raincoat in the bathtub, because it was still wet'

(raincoat wet? bathtub wet?)

SITUATION HOLISM

Eg Khoja restores the moon.

"One evening, Khoja looked down into a well, and was startled to find the moon shining up at him. It won't help anyone down there, he thought, and he quickly fetched a hook on a rope. But when he threw it in, the hook snagged on a hidden rock. Khoja pulled and pulled and pulled. Then suddenly it broke loose, and he went right on his back with a thump. From where he lay, however, he could see the moon, finally back where it belonged - and he was proud of the good job he had done."

Lycan, p.664

EXISTENTIAL HOLISM

This is the most profound phenomenon. Haugeland expresses it by saying: computers don't give a damn.

Computers have no sense of self, and so no sense of the self being diminished, or enlarged by experiences.

He gives some examples of where it is a sense of the sense being affected in these ways - through guilt, or shame, or embarrassment - that gives what happens its sense, and says that only by sharing such an awareness of self and its vulnerabilities can we understand some of the things that are said.

The distinction between cognition and affect - our emotional side and our apparatus for processing information - cannot be maintained when it comes to understanding natural language.

To make this point he gives an example and says that to translate an account of it into German one needs the translator to understand the embarrassing nature of the incident.

The case of the frghtened pet rat

' A friend of mine tells the story of the time she kept a white rat as a pet. It was usually tame enough to follow at her heels around the campus; but one day, frightened by a dog, it ran so far up her pantleg that any movement might have crushed it. So, very sheepishly, she let down her jeans, pulled out her quivering rodent, and won a round of applause from delighted passers-by.' Haugeland, in Lycan, p.667.

Haugeland says that most people find this anecdote amusing: and asks: Why?

Most of the amusement arises he says out of her embarrassment.

But embarrassment can only be experienced by a being that has some experience of itself - 'a sense that is important to it and can be awkwardly compromised on occasion.' p.667.

Only someone who could hear an account of that story and know that it was amusing could translate it successfully into German, say. German has several equivalents of 'sheepish' - with connotations of either (a) being simple; or

(b) being stupid; or

(c) being bashful.

'Only by appreciating the embarrassing nature of the incident could a translator make the right choice.

Moral: only a sense of self can give you a correct understanding of embarrassment, and computers would have to be given such a sense if they were to speak a natural language.

Another example.

Again, the point is that you need a sense of self in order to understand some of the things the tales that are told, and some of the things that happen.

Lucifer

'Ralph asks his new friend Lucifer: 'Why, when you are so brilliant, beautiful, and everything, did you ever get kicked out of heaven?' Rather than answer right away, Lucifer suggested a little game: "I'll sit up here on this rock," he said, " and you just carry on with all that wonderful praise you were giving me." Well, Ralph went along, but as the hours passed, it began to get boring; so, finally, he said: "Look, why don't we add some variety to this game, say, by taking turns?" "Ahh," Lucifer sighed, "that's all I said, that's all I said."

Haugeland, in Lycan, p.667.

Again: what is the joke?

Again: it is partly to do with this: Lucifer tries to pass off his cataclysmic betrayal as a little faux pas that anyone could slip into.

We only understand this because we ourselves have experience of guilt and shame (on a smaller scale). We know the temptation to escape from the guilt we feel for something by passing it off as something little.

But guilt and shame can only be understood in the way that counts by a being with some sense of self, some sense of how that self might be damaged.

Another example where 'the reader's personal involvement is vital'

Aesop's fable: loosing a son

Aesop:

One day, a farmer's son accidentally stepped on a snake, and was fatally bitten. Enraged, the father chased the snake with an axe, and managed to cut off its tail. Whereupon, the snake nearly ruined the farm by biting all the animals. Well, the farmer thought it over, and finally took the snake some sweetmeats, and said: "I can understand your anger, and surely you can understand mine. But now that we are even, let's forget and be friends again.' "No, no,' said the snake, "take away your gifts. You can never forget your dead son, nor I my missing tail.'

Lycan, p.668.

Everyday natural language is sophisticated in these ways too.

Haugeland challenges the conventional response to difficulties of this kind which suggests that understanding difficult cases such as these - understanding literature, as it were - can be left till later. For the time being we can address the simpler problems of understanding 'ordinary' English. Achieving that will be marvellous: and time to look beyond to the more sophisticated, more difficult problem of understanding more sophisticated forms of natural language.

Haugeland suggests that this is probably completely unfounded, and that 'everyday' language may be every bit as sophisticated as that used in more 'literary' forms.

Envoi

'Considering the progress and prospects of AI can be a peculiarly concrete and powerful way of thinking about our own spiritual nature.' Haugeland, in Lycan, p.669.