Machine translation (MT) is evolving steadily these days, based on the way artificial “intelligence” (I have to keep using these scare quotes to remind people how artificial all of this stuff is) is being developed, especially in the field of “chatbots” (artificial intelligence robots that we can chat with),*but I think some general points about it continue to be true even with this progress.
I think one of these is can be explained by comparing this subject to the old, old philosophical problem of “other minds.” Put simply, that is the question of how one individual person knows that others have minds essentially like theirs. All of us know this proposition to be true, it would seem; only extreme solipsists seriously doubt it. But why are we so certain about it, given that no one is “directly” aware of other persons’ intellects or consciousness the way we are of “our own”? All we seem to know without making inferences are the facts we perceive about another person’s physical behavior: what we see, hear, feel, etc., about them: all appearances of their bodies, nothing about minds.
Philosophers have made quite a few suggestions about how these inferences can be justified when they are carefully analyzed, including the view that inferring is not done at all in this area: we just know that other humans are fully human. As a graduate philosophy student, I actually devoted my Ph. D. thesis, long ago, to this subject, and like most philosophy problems, it never seems to get solved to everyone’s satisfaction. But here I want to consider how it is related to the subject of MT.
Suppose for a moment that one individual person, like you or I, is a solipsist, someone who refuses to agree that these other bodies that show up in their experience are equipped with these dubious things called “minds.” Those of us who are not solipsists would probably protest that the solipsist is merely scratching the surface, it might be said. How does the solipsist explain all of this bodily activity without philosophers and psychologists call a theory of mind, a theory that holds that the sounds emitted by those bodies, the “expressions” on their faces that seem so clearly to be expressions of emotions, and the movements of their arms, legs, and other body parts that one would assume to be obviously directed by a mind that wants to jump for joy, hang its head in shame, or just go out for a walk, are not mindless robot-like behavior?
We know that the work of a human translator (now dropping the solipsist philosophy) is the product of a mind that understands the meaning of the source text being worked on by the translator and is trying to re-express that meaning in a text in the target language. What connects these texts, one hopes, is the meaning. And what explains the connection is the translator’s grasp of that meaning.
In the case of MT, on the other hand, the so-called “neurons” in the computer systems are not anywhere close to as complex as the neurons of an animal’s nervous system, and the eighty or so billion neurons of the human brain far outstrip the “neural networks” of AI systems. This may account for the fact that young humans learn our languages with much less training than those systems need to handle only a small part of these languages.
Thus computers, even running the most advanced AI programming, don’t have anything like human understanding of meaning. (Even the most enthusiastic AI champions, I assume, would agree with this.) Instead, what the machine is doing is putting together a pattern of character strings (“ones” and “zeros,” or actually, high and low voltages in its circuits) by using some sort of very complicated circuit connections, which even the computer experts who create these programs tell us they can’t figure out, and drawing on the results of compiling millions or billions of pairs of texts in the two languages that have already been translated, probably by human translators, and stored in the computer’s memory. (Although in reality some of these “training data,” as they are called, were probably created by earlier machine translation operations, which raises the possibility of a “garbage-in-garbage-out” problem – faulty machine translations guiding later ones – especially since more and more machine-translated materials are fed onto the internet.)
It is this lack of an understanding of the meanings embedded in language that probably explains the shortcomings of machine translations that we see today and probably will continue to see for a very long time. This is the opinion of many AI experts, including Melanie Mitchell, whom I referred to in my previous post (in her book Artificial Intelligence: A Guide for Thinking Humans). Until computers are gifted with “minds,” human translators will still need to be called on, at least once in a while. For all we know, the longing that the positronic brain of poor Lieutenant Commander Data of Star Trek struggled with may still exist in the 24thcentury. And by then, of course, the Starship Enterprise will have a computer on board that an act as an interpreter for every language in the galaxy.
*For more information on chatbots, see Caldarini, Jaf, and McGarry: “A Literature Survey of Recent Advances in Chatbots” (https://www.mdpi.com/2078-2489/13/1/41). And you might want to check out the Wikipedia article on chatbots (https://en.wikipedia.org/wiki/Chatbot). And an article in the New York Times: “A.I. Is Not Sentient. Why Do People Say It Is?” (https://www.nytimes.com/2022/08/05/technology/ai-sentient-google.html). A fairly detailed general account of AI and how it can be improved is Gary Marcus and Earnest Davis: Rebooting AI: Building Artificial Intelligence We Can Trust (Pantheon Books, 2019). See also Erik J. Larson: The Myth of Artificial Intelligence: Why Computers Can't Think the Way We Do (Belknap Press of Harvard University Press, 2021).
コメント