« Iranian missile hits Israeli ship | Main | The salary taboo »

July 18, 2006

Comments

I always thought that a robot translator (or something that could effectively teach someone a foreign language very quickly) would be a bigger threat to menial labor workers. Because then immigrants could work cash registers, telephones, etc.

All the assembly jobs moved to China years ago anyways.

Actually, I'd say that those who can't figure out a bookcase are those whose skills lie elsewhere, like executives, doctors, etc.

I'd agree with Dennis.
Also, a robotic assembly line could always assemble a particular bookcase.
The reason that Ikea bookcases don't come assembled is for easy transportation. Until everyone has their own personal home robotic servant robots that can assemble such cases won't be of much help. I expect surgeons to be replaced far earlier.

Someone who is doing translation is pretty much by definition not a menial worker. Though I see that your point is really about eliminating one of the obstacles to the employment of immigrants. Having worked on a speech-to-speech translation project (project CSTAR at CMU), I can assure you that translation is 'AI-hard', that is, you won't have good translation until you can create general human-level intelligences or something close to it.

Bbartlog, I'd love to see you put your money on the assertion that translation is "AI-hard". Major AI researchers (like Hoffsteader) once said the same of champion level chess and I strongly believe in information markets as a method of aggregating information when people disagree on testible matters of fact.

Chess is only an 8 x 8 grid, and through brute force the computer can see more moves ahead than a human.

Computer programs can't even win at poker against expert level players (run of the mill experts, and not super-human people like chess player Gary Kasparov).

People like to joke about how klutzy they are, but I've no doubt that the average person CAN assemble a bookcase (or just about any other item) from IKEA. People who claim they can't usually just don't have the time or patience.
No doubt some people have physical impediments such as poor dexterity or eyesight, but that's surely the exception.

What would constitute a test of translation? Also, do the online betting intermediaries force you to put money in escrow to make a bet? If so I'd be sacrificing returns for quite some time... in principle I would be willing to make such a bet however, depending on the nature of the translation task.

Two little examples from my work of 1999-2000, which don't even touch on really hard stuff like metaphor and poetry:

Computers are lousy at dealing with anaphora. This is the use of words like 'it' or 'that' to refer to things previously mentioned or obvious from context. Given a large enough corpus, statistical methods will provide means for dealing with frequently-occuring phrase (like 'You want fries with that?' or 'It's raining hard'). But it's not hard to construct a conversation littered with anaphora that will boggle a computer. Note that language differences make it impossible to just 'pass through' the anaphora in hopes of the other speaker decoding it.

Second is the problem of dealing with negation. The questions 'Do you want fries with that?' and 'Don't you want fries with that?' appear, based on literal interpretation, to be opposite in meaning, so that answering 'Yes' to the second question might appear to indicate 'yes, I don't want fries' (but in common English usage of course means 'yes, I do want fries'). In German, the word 'doch' provides an interesting answer to such questions (sort of equivalent to 'actually, yes' in this context). However, not all languages allow for this type of question, where a seeming negation of the question allows for the same set of answers as the positive formulation. A computer has to be smart enough to either strip the negation from the original question (losing nuance, in this case, as the two question don't really have fully identical meaning), or else know enough about the exchange to flip the affirmations and negations of the speakers as appropriate. This is hard.

For computers playing poker, the simple case of heads-up limit hold'em is nearly solved. I agree that both no-limit and many-handed poker present special challenges that make a general solution very difficult. Also, it's fair to ask whether you want the computer to play an unbeatable game (in a game theoretic sense, i.e. a game with no weaknesses), or whether it should play to also exploit weaknesses of its opponents as a human expert would.

Allright, forget translation. Let's imagine that someone actually invents an interactive game that a PC or a Sony Playstation can use. This game is fun and interesting and forces you to practice English. Maybe after 6 months or a year of regular play, you have mastered conversational english. That would allow immigrants to compete for a giant set of low-wage careers they are currently excluded from due to their language limitations.

People are already learning how to speak English. In the Phillipines, being able to speak English guaranteed the speaker a job at a call center that pays a big premium over local wages, but of course that big premium is still less than minimum wage in the U.S.

On a related note, an Economist article on a robot to make clothes: http://economist.com/science/displaystory.cfm?story_id=7159425

"People who claim they can't usually just don't have the time or patience."
(Though I don't usually like anecdotes...)That about describes my father. I figure as you do that he could, say assemble a bookshelf, but he makes enough that it's irrational to do so.

The comments to this entry are closed.