next up previous contents
Next: Experimental Possibilities Up: THE WELL-DESIGNED CHILD Previous: What Abilities Could Usefully

Features of a Language of Thought

Cognitive scientists argue about whether there is a language of thought, but its advocates haven't told us much about what it is like. Stephen Pinker, an advocate, only tells us in [Pin94]

The hypothetical ``language of thought'', or representation of concepts and propositions in the brain in which ideas, including the meanings of words and sentences, are represented.

A language of thought that might be used for robots or looked for in humans is constrained by the characteristics of the baby's world and the characteristics of the non-linguistic parts of the baby's mind--including its limitations.

Here are some ideas about mentalese.

grammar is secondary
While most linguistic studies have focussed on grammar, meaning is more important--in studying spoken language, in proposing a language of thought and in designing robots. A child's first speech consists of words which are attached to things, or to appearances of things, or to sometimes ambiguous combinations of things and appearances. ``Doggie'' is stimulated by the sight of a dog, a picture, an animal on TV, the sound of barking and conversation about dogs.

maybe language starts with naming
A human child starts language learning with names for objects. This desire is independent of having any goal concerning the object. We have the option of designing an artificial child to know a lot of language, e.g. English and/or a logical language from the beginning. Different experimenters will explore different approaches.

parallel information
Images are presumably represented in parallel. There is nowhere anything like a television signal processor that handles a picture serially and repeatedly spreads it out. This is obvious for pictures but surely applies to a lot of other kinds of information. On the other hand, our inability to think completely in parallel shows that many higher mental functions are done serially.

For a robot, a logical language 12 will be most suitable, but some appropriate ascriptions of beliefs and intentions to robots will refer to information represented non-logically. Humans probably don't use quantificational logic at the pre-verbal level, although we can use it when we have to, and formal logic is often helpful when the information is mathematical. Here's why I only say probably. Consider the sentence ``For every boy there's a girl who loves only him.'' Its predicate calculus representation, , has three embedded quantifiers. We then ask the question, ``What can you say about the total number of boys and girls?'' A fair number of people uneducated in logic find the correct answer that there must be at least as many girls as boys. I haven't done the experiment thoroughly, but the results suggest that the sentence with the three levels of quantifiers is understandable by many logically uneducated people and therefore its content is somehow internally represented.

Let's design it into our child.

a word at a time
Sentences uttered by humans are usually not preformed in entirety before being uttered. A human starts a sentence and thinks how to continue and finish it as he continues talking. (The obvious argument is from introspection, but I suppose experiments would confirm it.) Humans can preform sentences with some effort. Vladimir Bukovsky tells about having composed a whole book in prison while denied paper.

chemical state
Suppose a person is hungry--a condition humans share with dogs. This can perfectly well be only be represented by the chemical state of the blood stream. There is no reason to have anything like the sentence, ``I am hungry'' anywhere in the brain until the fact has to be communicated. Similarly we don't need anything like a sentence in the memory of a computer to represent the voltage of its battery.

virtual sentences
We may regard information that is directly represented by the chemical state of the bloodstream or by a voltage as expressed in virtual sentences along the lines of [McC79] or [New82]. We may then sometimes be able to explain some actions as involving logical inference involving the virtual sentences.

immediate reference
Thinking about an object before one's eyes does not require that it have a name. Something like a pointer to a structure will do as well. We can see this, because when we have to mention an object in speech we have to think of a name that will enable the hearer to establish his own pointer to his mental structure representing the object in question. Purely internal symbolic names as in Fodor's proposed language of thought may be useful even if they aren't communicable.

short thoughts
Thoughts are not like long sentences, although a long sentence may be required to express a thought to another person because of a need to translate internal pointers into descriptions.

When the fact of hunger or low battery voltage has to be communicated something like a sentence is needed. Let's call it a pseudo-sentence until we find out more. However, a pseudo-sentence isn't needed to stimulate eating. It also isn't necessary to represent the rule, ``if hungry, then eat''. In view of evolution, one would expect the fact of being hungry to be represented both chemically and in the language of thought.

There are other uses besides communication for sentence-like forms. Very likely, the expectation of being hungry by dinner time needs something different from a substance in the blood for its representation.

The language of thought is used for reasoning.

not like spoken languages
English and other spoken languages won't do as languages of thought. Here are some reasons. I think there are additional reasons, but I haven't been able to formulate them.

The language of thought may undergo major reorganizations. This may be one reason why there is so little memory of early life. Almost no-one can remember nursing or drinking from a baby bottle.

next up previous contents
Next: Experimental Possibilities Up: THE WELL-DESIGNED CHILD Previous: What Abilities Could Usefully   Contents
John McCarthy