AI: The Technological Trickster
By ai-depot | October 26, 2002
The Language of Thought
Our ability to communicate and manipulate our environment using spoken and, more recently, written language, is one of the most important areas of brain and computer research. Again we find similarities and differences, but it is still unclear whether the differences are of magnitude or of inherent inabilities to translate from one medium to the other. Much depends on our definitions of thinking and how that relates to language development. “The essential element of linguistic competence is a (shared) representation that is general…and is extendable…” (Fischler and Firschein, p. 185) Language must use symbols that everyone can recognize and manipulate within the confines of the particular linguistic structure. And it must be extendable potentially to infinity in order to deal with changing reality and to incorporate new knowledge.
Several lines of research have led to better understanding of how we learn language. From studies of children in the process of learning speech, the evolution of speech in humans, and attempts to teach other primates to speak, to the current information processing of computers and how that may reflect our own abilities and limitations.
Noam Chomsky says “Language can be used to transmit information,…to establish relations…, to express or clarify thought, for play, for creative mental activity, to gain understanding,…” (Chomsky, 1977. p. 88) and so forth. For AI researchers the problem becomes how to inculcate a computer with the ability to interpret what it is doing when it is forming thoughts, if it is forming thoughts, and how it may manipulate symbols in this regard.
Chomsky’s hypothesizes that humans are born with a “language acquisition device” (LAD) which seems to act as a little computer that incorporates a bunch of general rules of language and manipulates them so as to develop an understanding of the particular language of its society. Attempts to localize or to discover how such a system might work in the brain have not been particularly fruitful.
On the other hand, Terry Sejnowski and Charles Rosenberg used a computer to simulate a three-layer motorneuron network with specialties for phoneme pronunciation. While no one site was specialized for any one phoneme, through random interactions the group was able to reach a ‘babbling’ stage of pronunciation in a short time, eventually learning a complete sentence, without the input of any rules of grammar. (Calvin, p. 233-235) This suggests again that the whole brain plays a part in the learning of a language, and that the complexity of the system is able to encompass this ability. Again, we will likely find selection pressures behind the ability to develop and utilize language.
While “…Chomsky… shares with Shannon an inclination to separate form from meaning, and let the study of form reveal new truths…” he also “…makes a critical distinction [in language use] between competence, which is regular and orderly, and performance, which he sees as entropic, noisy…” (Campbell, p. 161, 168, 185) Performance may be seen at all levels of language learning, but competence comes with a decrease in entropy that may be the result of repeated use. Yet Chomsky’s work indicates there is much more than simple repetition involved in developing competence. Language is both rule-bound and potentially infinite in its ability to express. A sentence can incorporate as many “ands” and clauses as necessary. This process may not begin with rules, in fact may be antithetical to any rule-based system.
Lenneberg’s studies of child learning suggest “…a fundamental principle of language acquisition: what is acquired are patterns and structure, not constituent elements…” (Lenneberg, p. 79, 218) We may be able to proceed directly from patterns to meanings if we are part of a group that agrees on the basic premises. If so, it seems particular symbols may be added to the system after social agreement has been reached on how a concept or state is to be expressed.
Gardner and Winner, in “The Development of Metaphoric Competence” find that children metaphor at an early age. The youngest children tested produce the highest number of both appropriate and inappropriate metaphors in a test that asks them to finish the sentence, “He looks as gigantic as…” While observing that context is essential in interpretation (and presumably in the formation) of metaphors, the experimenters interpret this to indicate an “…insensitivity to…conventional boundaries of experience and language.” (Sacks, p. 138) It can also indicate children are more apt to play with their new found abilities to describe their thoughts, and that their metaphors fail more often than older children’s metaphors.
These observations indicate the ability to metaphor, to allow one sound, thing or symbol to stand for another, is important in language acquisition. The appropriateness of a metaphor depends on context, and in a sense, all the children’s metaphors were adequate to some situational context. (Although their similes may not have been) Douglas Porteous believes “…all thought may be, at base, metaphorical and…language develops by way of metaphor.” (p. 69) In fact, to impart meaning to a sound is in itself metaphorical.
Some researchers suggest a less prominent role for the ability to metaphor. “Analogies and metaphors simply point the way to where useful blueprints might be found.” (italics mine)(Waldrop, p. 57) This sentence contains an interesting use of at least two metaphors to underestimate the power of metaphor.
Psychologist Julian Jaynes suggests that in the evolution of language, “…language and its referents climbed up from the concrete to the abstract on the steps of metaphor…” and that “The lexicon of language …is a finite set of terms that by metaphor is able to stretch out over an infinite set of circumstances, even to creating new circumstances thereby.” (Jaynes, p. 51, 52) It seems this ability to evaluate one thing in the light of another may be at the root of language acquisition, and may be basic to a number of brain activities. The ability to reduce a complicated problem with no apparent solution to a less complicated one with a known solution probably has a good deal of survival value, among other useful traits. And it suggests as well that meaning is attached to symbol rather than the reverse.
John Eccles has noted that “… though the monkey and the ape have …all the neural machinery for cognitive and motor learning, they are greatly handicapped in a novel situation by being unable to think… linguistically.” Primate researchers Sasaka and Gemba, using a light stimulus to activate a reward, note that even an intelligent monkey may require 4000 trials for a task a human preschool child learns in only a few trials after verbalizing it. (Eccles, p. 166)
Eccles further notes Geschwind has observed that in the human brain “…areas 39 and 40 (Wernecki’s area) are the only new structures that have appeared in the evolution of the human brain…[and] that area 39 (the anular gyrus) has a special function in speech since it developed at the meeting place of visual and tactile information…” (Eccles, p. 92) The ears and tongue are left free to symbolize independently of this system, suggesting feedback and feed-forward mechanisms in language development.
Further, Kent agrees with “…Liberman’s idea that the functional specialization of the language hemisphere is that of making the acoustic to the phonetic level of translation.” (Kent, p. 220) Humans may be quite a step beyond chimpanzees in this regard, but the gap between humans and computers must be very much larger.
Eccles suggests there have been four levels of language evolution, (developed through his work with Buhler and Popper), which he calls the expressive, signal, description and argumentative stages. (Eccles, p. 71-72) Computers may be able to incorporate the first two, and we can see how they may be able to learn to describe, but the last stage is at present beyond the capacities of any known computer program. Yet with careful programming an illusion of argumentation is possible in a computer, as Joseph Weizenbaum discovered in the reaction to his ELIZA program which he based on Rogerian techniques of psychoanalysis. By rephrasing a patient’s statements into appropriate questions about further mental states, some degree of knowledge is brought to the surface, but not necessarily in the machine.
The problems of generalizing or metaphoring from human language to computer languages are complex. Computer languages seem to fit all the criteria of language, yet there are problems in how these programs fit with the world. Weizenbaum’s ELIZA (named after Shaw’s Eliza Doolittle of Pygmalion who, although she adopted the speech mannerisms of the British aristocracy, it was not clear that her thinking had changed) has illustrated the problem, “…to construct a finite program that assigns appropriate conceptual structures to the infinite range of sentences that can occur in natural language… remains as untouched as ever.” (Weizenbaum, p.199)
While some strides have been made since 1976, the present state of affairs seems to range from cautious optimism to dangerous misconceptions. Computer translation is not able to deal well with the changing nature of language and languages. One of the difficulties with having a universal language such as Latin or Esperanto is there isn’t the ability of these to fluxuate and change with use as the world does, which mirrors the computer problem.
Weizenbaum suggests computer “…systems and programs… are all…simple…[that they] distort and abuse language… and…while disclaiming normative content, advocate an authoritarianism based on expertise.” He warns that the “… computer has become…an instrument for the destruction of history.” (Weizenbaum, p.238, 248) The inappropriate metaphor may put us in a bind from which it is difficult to extract ourselves in much the same way inappropriate metaphors have resulted in conflicts between cultures. Language is context specific in a number of ways that indicate “…human language behaviour is part of a coherent plan of action…” (Fischler and Firschein, p. 183) We talk about things with which we are intimately involved. The danger in divorcing language from meaning in order to study it may distort the object of study beyond our ability to put it back together in a meaningful way.
We must be cautious about the mistake the Raven made lest we lose our ability to verbalize completely.
Tags: none
Category: essay |
