Interviews

I have been conducting a series of interviews over the last few weeks with my Turkish associate (now friend) Ugur Polat. We have done four so far, each over three hours long (using Zoom). He plans ten in total. The actual interviewing is done by his neurologist wife Burcu, because her spoken English is better (very good in fact). We have been covering my intellectual life in detail. So far, we have dealt with my years in Manchester, London, and Oxford. The questions are probing and well-informed; I get to talk about my old friends and colleagues, many now deceased. He is planning to make a book of it to be published in Turkish and English. He is also interested in writing a book about the hand, starting with Sir Charles Bell and ending with me—an excellent project. In addition, I treated them to a mini-concert of songs in which a good time was had by all.[1] I urge all rabid feminist cancellers (bless their vicious little hearts) to contact Ugur and try to prevent him and his wife from carrying out their intentions—good luck with that! It’s nice to be in contact with decent people.

[1] I performed Black is Black, You Better Move On, Everything I Own, The First Cut is the Deepest, Leave my Kitten Alone, Stand by Me, Love Hurts, and Alone.

Buying a Language

I want to learn a language, say Spanish, but I can’t be bothered to do it the old-fashioned way, so I decide to buy it (expensive but labor-saving). I go to the language store (Lang-Mart, right next to the Chinese supermarket). I am told that I can buy either a grammar chip or a dictionary chip or both. I figure I will buy the grammar chip and use an English-Spanish dictionary for the word meanings—cheaper that way. I go home and insert the new chip into my brain: by pressing a button a whole lot of Spanish words crowd my consciousness; lo and behold, strings of them form in response to the environment and the utterances of Spanish speakers. I proceed to look up these words in my dictionary. But this is slow work and it quickly becomes apparent that it is impracticable; as expected, I don’t know the meanings of Spanish words. This is no use, I think, and return to the store. I swap the grammar chip for a dictionary chip that will install knowledge of what Spanish words mean in my memory, hoping that I can figure out the grammar myself. I go home and install the new chip: miraculously it gives me complete lexical knowledge of the Spanish language. But the grammar defeats me; I don’t know how to combine the words into grammatical strings (I have never been good with grammar). I trudge back to the store and reluctantly buy the grammar chip as well (pricey!). Now I am in business: I can now form grammatical strings that I understand. A language consists of a grammar and a lexicon, so both things have to be known in order to speak the language like a native: atomic knowledge of word meaning and combinatory knowledge of grammar working together.

If a machine (call it a computer) can only do the combinatory part, it won’t know the language properly, just a component of it. If another machine (call it a mental dictionary) can only do the atomic part, it also won’t know the language properly. Thus, the existence of either type of machine will not show that a machine could speak in the normal way: symbol manipulation is not enough, and lexical information (knowledge of word meaning) is not enough. Language mastery needs both. Accordingly, there can be a Chinese room argument against computers as competent speakers, but there can also be a “Spanish room” argument against machines that have lexical data banks (word meaning memories) but no grammatical competence. We would need a machine that combines both to have a genuine speaker-hearer. We actually have machines that do the latter (albeit without real grammatical knowledge), but we don’t (yet) have machines that contain the latter (knowledge of word meaning). In order to know what “red” means, say, a machine would have to be consciously acquainted with redness: but no machine yet built is (as far as anyone knows).[1]

Would a machine that has both types of knowedge know a language like a normal human? I’m afraid not, because in addition to syntactic and semantic knowledge a competent speaker needs pragmatic knowledge. Suppose I want to learn to speak Japanese like a native: I go the language store to pick up the necessary chips only to be told that I will need a pragmatics chip as well as the other two. Unfortunately, this one a really quite expensive—$2,000 a pop! But if I don’t have it, I won’t be able to use the language in ordinary conversation, because I won’t know the rules of appropriateness and politeness that govern social language use. I could get arrested if I say the wrong thing! I fork over the dough. Thus, there could be a “Japanese room” argument to the effect that a machine could never acquire pragmatic knowledge of a human language, even it could have syntactic and semantic knowledge. You might be fed syntactic and semantic information and be fully competent in those aspects of language, but this information omits the pragmatic aspects of speech, which require knowledge of culture, human psychology, etiquette, historical context. None of this is entailed by syntactic and semantic knowledge, which is machine-like in comparison to pragmatic competence. Pragmatics can’t be reduced to syntax and semantics, as semantics can’t be reduced to syntax; so, a machine that can do only syntax and semantics will not be a fully competent speaker. A machine that can only do phonetics will likewise not add up to a speaking machine, since syntax goes beyond phonetics; just because a machine can make the sounds of English it doesn’t follow that it can speak according to English syntax. The so-called Chinese room argument is one in a series of parallel arguments directed against claims that specifiable types of machine really have language mastery in the normal sense. It says, essentially, that there is more to semantics than syntax (grammar)[2]; but we could also argue that there is more to syntax than phonetics and more to pragmatics than semantics (and more to poetics than pragmatics). All true, but one wonders why any elaborate argument was needed in the first place. Speaking a language is a complex skill consisting of several components that can be independently possessed.

[1] It is clearly not enough to have a machine containing a dictionary, since it would need to know what the defining words mean. At a minimum this requires knowing what those words refer to. Translation knowledge is not the same as semantic knowledge.

[2] The only way to resist the argument is by claiming that computers do more than manipulate symbols; they also refer to things with those symbols. They do so by being embedded in the world in such a way that reference naturally follows, as the causal theory of refence maintains. Then the question will be whether this theory is true. I myself think that computers are nowhere near mastering a language as humans do, but there is no reason of principle why a machine can’t speak like a human (just build one that exactly resembles a human down to the last detail and then rely on supervenience).

Chinese Rooms and Linguistic Knowledge

John Searle’s Chinese room argument shows that it’s possible to be able to form meaningful sentences in a foreign language without knowing that language. That is, you can know the grammatical rules of a language without knowing the lexicon—what the individual words mean and refer to. Suppose you wanted to learn Chinese and you go to an instructor offering to teach you that language, but all he does is tell you how to put words together into meaningful strings, omitting to tell you what the individual words mean; you would not learn the language—at best you would learn its grammar. Your teacher might assign each word to a grammatical category, so that you would know which are the nouns and which the verbs, etc.; still, you would not know what the words mean. Surely, this is trivially true: knowledge of grammatical category and rules of combination do not suffice for full knowledge of the language. Nor will having consciousness help, or even conscious thought; you still don’t know what those alien sounds and shapes individually mean. If you did, you would know the language—understand it—but you don’t. Knowledge of language consists of two parts—knowledge of grammar (syntax) and knowledge of word meaning (lexicon)—and the former does not entail the latter.

Does this show that a computer doesn’t or can’t understand a language? That depends on what is true of the computer. It doesn’t if it has no knowledge of what the words it manipulates mean; but if it contains that knowledge, it does know the language. It’s the same for a conscious human being: if he lacks lexical knowledge, he lacks knowledge of the language, no matter what else is true of him. The question then becomes, “What does it take to know the meaning of words?” This is where the problems begin: for what does it take for me to know what words mean? What does it take for anyone to know anything? If we knew that, we could decide whether a computer (what kind of computer?) can know and understand a language. There are many theories of linguistic knowledge: imagistic theories, use theories, intention-based theories, causal theories, neural network theories, description theories, indeterminacy theories, community practice theories, etc. None of these commands universal assent. It’s a philosophical problem (also scientific). In order to get a machine to understand a language, you would need to install whatever conditions your favorite theory prescribes; and that will be a substantive question. Can a machine be programmed to have mental images or intentions or agency or “forms of life”? The Chinese room argument doesn’t say; all it says is that grammar is not enough. And it may be that a computer can’t know or understand grammar either—it merely simulates such knowledge. Adding consciousness will make no difference, since many animals are conscious without being able to understand a human language. The Chinese room argument is silent on such questions, and irrelevant to them. What it tells us is true enough, but it doesn’t settle the question of machine understanding. I would say that we don’t know what gives us lexical knowledge, as we don’t know what gives us consciousness (and other mental capacities), so we don’t know whether a machine (whatever that is) could have knowledge of language. It may or may not be the same as what makes a biological organism linguistically competent—we just don’t know. To put it more polemically, the Chinese room argument is toothless; it simply shows that grammatical rules don’t entail lexical information—which is trivially true. I suppose we could say that it shows that our current computers don’t understand language because all they can do is manipulate symbols whose meaning they don’t know. This might well be said of some conscious and intelligent biological organism that happens to be good at grammar but bad at words; all it can do is string words together grammatically without having any knowledge of what those words mean (some sort of idiot savant might be like that). I don’t think any computing machine we have invented so far really knows a language as humans do, but that doesn’t rule out future machines that do better. At a minimum they would need conscious perception so as to ground object-directed thought, but we know so little about perception and cognition that we can’t say whether we will ever build such machines. All Searle’s argument shows is that (mechanical) syntactic engines fall short of semantic understanding, but that was surely obvious all along and doesn’t require the creation of an elaborate thought experiment. The argument might be re-labeled the Syntactic Impotence argument. This says that you can arrange words into patterns (like jigsaws) without knowing what they mean. Yep, so what?[1]

[1] It might be said of computational enthusiasts that they can’t see this point, obvious as it is, so it’s worth arguing with them. Maybe, but the argument is very limited on the substantive question at issue, viz. whether a machine could think and understand language. My own view is that we are machines, so the answer is yes. As to knowledge of word meaning, I am a mysterian (like Chomsky).