There is no single correct interpretation of an image except, perhaps, within a domain. For
example, on a Forest Service map, the oval might unambiguously mean a campsite and a small
rectangle might mean that the campsite has a picnic table. In another domain, the same symbols
may mean something entirely different. There is no well-defined limit to the number of possible
domains in which an image could have particular domain-specific meaning and no limit to its
possible interpretations in general language. The world is infinitely categorizable. But
interpretations are not exactly random meanings either. Each is somehow motivated by the
original image. The first interpretation of the image in my dream is motivated by a similarity
with the shape of a balloon filled with water. The second suggests the unjustified
oversimplification of a human to a stereotype which supposedly allows us to predict how that
person will behave (a kind of extreme racism). And the third interpretation of the line and oval
suggested at a second level the story of the beam of light shining through a knot-hole into a
tree house. One could say that the first interpretation is literal while the second is metaphorical
and the third is both literal and metaphorical. But note that, contrary to a common assumption
about metaphor, the metaphorical interpretations are not based on the literal balloon
interpretation. ^{(note 1)} However, the beam-of-light story has a particular significance to the
philosophy of language related to the second metaphorical interpretation. We concur with the
philosopher Emmanuel Levinas who claims that selfhood is based on a recognition of otherness
in the sense that other people also possess selfhood. Without others, selfhood has no meaning.
However, with particular individuals, we can attempt to deny them their agency. One way is to
stereotype another person and claim that a label is all we need to know about the other (as in
the second interpretation of the oval). Another way to attempt to deny the agency of others is
to pretend that they are not relevant to us by putting up barriers around us to shut out
everything but our own little world (as in the story of the tree house). When we stop resisting
the otherness of others and grant them the same agency we possess, then a whole new world
opens up (suggested by looking along the beam instead of at it), an exciting world which is
fundamentally ambiguous yet grounded in the ethics and economics of human relations, a
world which opens up the possibility of dynamic general language.

As long as we are comparing general language to a world, let us extend the metaphor to include
domain-specific language. Start with our planet earth and its various languages/cultures focused in
various geographical areas; then think of the various satellites orbiting the earth as artificially created
domains. Some domains, such as the domain of the maintenance and repair manuals for a piece of
machinery sold world-wide, will be almost completely shared across several languages, just as one
satellite can transmit to several areas of the earth. Even then there will be minor variations such as
the voltage and frequency a machine expects when fed electrical power and the type of plug placed
on the outlet to obtain that power. Other domains are tied to one culture, sometimes even within the
same language, just as a weather satellite and a military satellite may be in the same orbit yet be
incompatible. For example, the domain of Law in the United States and the domain of Law in Great
Britain are two domains, whether you count British English and US English as one language or
two. Translating between incompatible domains can be as challenging as translating between general
languages. ^{(note 2)}

Human translators are able to handle both general-language and domain-specific texts. As a
starting point, a translator must be competent in two or more general languages. Then, for each new
domain, the human translator must gain new expertise. The same requirement applies to a machine
translation system in that the lexicons as well as any knowledge base the system may have must be
updated in order to produce high-quality translations from a new domain. But here the similarity
ends. Human translators can produce high-quality translations of general-language texts which are
dynamic, that is, full of metaphor, allusions, and intentionally unusual usage. Current machine
translation systems cannot. Current techniques in machine translation produce fully-automatic high-
quality translation only when applied to a body of similar texts which are all restricted to the same
domain. The texts must be static in that they do not contain new metaphors, allusions, or
grammatical constructions. Sometimes this restriction occurs naturally and the texts form a
sublanguage. More often, the restrictions must be enforced with the cooperation of authors,
resulting in what is called controlled language. Many have noted that machine translation works
better in a narrow domain. The reason in obvious: everything is better-defined and less ambiguous
than in general language. What is less obvious is whether the machine translation techniques that
work quite well within a domain can gradually be extended to apply equally well to general
language. Or will one encounter a phenomenon of diminishing returns or even an unscaleable wall.
Terrance Hook, who has developed a domain-specific Dutch-English machine translation system,
made a typical comment. He said that when restricted to a domain, the output of his system is good
enough to be used, as it is, for some purposes. However, when, 'on a rainy afternoon', he tries a
passage from a newspaper, he gets gibberish. Is this a temporary limitation of domain-specific
systems or will they gradually improve in their ability to handle general language texts until they do
as well on general language as they do on domain-specific texts? We claim that current techniques
of machine translation will never be extended to handle general language texts.

Techniques that do not extend

A major shift has occurred in machine translation. John Hutchins, the acknowledged historian-
in-residence of machine translation, has noted that up to about ten years ago the assumption
was that systems should be general; ^{(note 3)} but now the assumption is that systems (at least systems
aimed at high-quality output) should be domain-specific. The issue among professionals is no
longer whether current techniques in machine translation work equally well in a domain and on
general text. They do not. The issue is whether current techniques can ever be extended to
handle general language effectively. I have proposed that they cannot be extended to dynamic
general language. This claim is highly controversial. ^{(note 4)} How could I be so bold as to make such
a claim? The reason is based on: (1) the fact that current techniques depend on philosophical
underpinnings called objectivism; and (2) my belief that general language does not conform to
the assumptions of objectivism, thus invalidating current techniques as applied to general
language.

George Lakoff, a prominent linguist and early supporter of Chomsky, long ago broke off from
the objectivist camp and has spent recent years developing a non-objectivist approach called
experientialism. He summarizes objectivism as the belief that:

rational thought consists in the manipulation of abstract symbols and that these symbols
get their meaning via a correspondence with the world, objectively construed, that is,
independent of any organism           (Lakoff 1987: xii)

This view has many implications. It implies that the human mind is an abstract machine and
that any machine, including a digital computer, which is properly programmed, is theoretically
capable of thinking just as well as or even better than a human mind. Note that this view
includes a strong form of mind/body dualism, which means that a human body is not at all
necessary for human-like thought. Some researchers in Artificial Intelligence (AI) take what is
known as the strong-AI position, which is that computers will someday be able to perform any
intellectual task that humans can perform. ^{(note 5)} Marvin Minsky, a strong-AI proponent, recently
wrote an article in which he estimates the knowledge a human acquires over a lifetime amounts to
not more than the equivalent of about three gigabytes, which is approximately the amount of
information that can be stored on one CD-ROM. He speaks of nanotechnology that places
individual atoms in desired positions and that will allow us to produce much smaller and faster
computer chips than we now can build. He then states, speaking of future robots as our virtual
offspring, our MIND-CHILDREN:

Once we know what we need to do, our nanotechnologies should enable us to construct
replacement bodies and brains that will not constrain us to work at the crawling pace of 'real
time'. The events in our computer chips already happen millions of times faster than those in
brain cells. Hence, we could design our 'mind-children' to think a million times faster than we
do. (Minsky 1994:90). (Scientific American, October 1994, pp. 86-91)

Minsky also notes that many scholars from a variety of disciplines 'firmly maintain that
machines will never have thoughts like ours because, no matter how we build them, they will
always lack some vital ingredient'. Minsky says he has no patience with such arguments
because they are all flawed by assuming, in one way or another, 'the existence of some magical
spark that has no detectable properties'.

Although over the years I have generally had little patience with Minsky and his outrageous
claims, he has a good point here. ^{(note 6)} In a post-religious society such as ours, it does little good
to use an 'undetectable magical spark' as the basis for an academic claim. Instead I have
decided to focus on what hurdles would have to be overcome by a machine before it would
even have a chance of handling dynamic general language better than or on a par with humans.
I do not claim that it will never be possible to build machines that can think like humans and, in
particular, can handle dynamic general language as well as humans. Instead, I try to show that
the current techniques of natural language processing (NLP) will never be extended to
accomplish such tasks. Entirely new techniques will be needed. In particular, we will need
techniques that avoid the assumptions of objectivism. We will see why in the next section.

Avoiding objectivism

Both mainstream philosophy and mainstream linguistics have built into them assumptions based
on objectivism. Here are some of those assumptions:
(a) Words and fixed expressions such as multi-word terms are mapped to a short list of discrete
senses, often to a single sense.
(b) Each sense exists independently of any particular word or sentence and has the properties of a
mathematical set. For example, the sense of horse that corresponds to an animal is a set of objects in
the real world. Any particular object is either in the set (if it is a horse) or is not in the set (if it is not
a horse). There is nothing in between. Since these senses are independent of particular sentences
and independent of people, they correspond to the way the world is, to the way the word
objectively divides itself up.
(c) The meaning of a sentence treated in isolation can be obtained by combining word senses of the
words of the sentence from the bottom up. If a word of the sentence is ambiguous then there may
be multiple composite sentences for the sentence, unless all but one are weeded out by selectional
restrictions.

These assumptions are embedded in the standard framework which divides language into
syntax (including morphology), semantics and pragmatics, with emphasis on syntax and
semantics. According to this framework, linguistics is a branch of individual rather than social
psychology. To someone committed to the mainstream view, this framework is perfectly
standard and obviously true. There are many flavours within Generative Grammar, but they all
share this framework and most work in machine translation is explicitly or implicitly based on
it.

However, dynamic general language violates all three of the basic assumptions listed as (a),
(b) and (c). It violates assumption (a) in that new word senses, sometimes called nuances, can
be generated dynamically as needed in speech or writing, often for the purposes of a single
text. Indeed, this dynamic aspect of meaning is found in all interesting writing, not just in great
literature. Only in a well-defined domain can the meanings of words be pinned down. And that
is because we humans create a domain specifically so that the senses of a term will be limited
and discrete, with the goal being one concept per term and one term per concept in each
language.

Dynamic general language also violates assumption (b) in that its categories are not
mathematical sets tied directly to the way the world divides itself up. Lakoff (1987) gives
abundant evidence to this effect from several disciplines. For example, he shows that categories
of general language exhibit prototype effects in which some members are better members than
others, a behaviour not allowed in mathematical sets. ^{(note 7)}
Again, in a domain, we divide up the
world a certain way for a particular purpose. So from the point of view of the domain, the
world can be seen as divided up into a neat ontology of domain concepts which are
mathematical sets.

Assumption (c) is violated in that general language is always understood in a certain context.
Martin Kay and his colleagues (Kay et al 1994) put it this way: 'language is situated'. When
humans process general language, they do not delay consideration of pragmatic factors such as
the situation. The syntax, then semantics, then pragmatics model is only applicable to domains
in which the situation is constant and therefore implicitly taken into account at all levels.

So we see that dynamic general language violates all three assumptions on which most
natural language processing is based. But controlled language restricted to a well-defined
domain conforms to all three assumptions if we engineer it so. At a dinner speech, Martin Kay
once put it something like this: 'Success in NLP has been seen primarily in cases where natural
language resembles formal language'. That comment, although intended to be humorous, is on-
target and has a serious side. The syntax/semantics/pragmatics model of bottom-up
composition from well-defined concepts is essentially a description of a formal language such
as a computer programming language. Formal languages conform to all three assumptions
while dynamic general language conforms to none of them. Thus, NLP techniques that are
based on these assumptions apply to domain-specific text inasmuch as it resembles formal
language and inasmuch as it does not exhibit the dynamic possibilities of general language.

Thus we can conclude that current NLP techniques will never be extended to handle dynamic
general language, since to do so they would at least have to abandon the three basic
assumptions of this section. Any set of techniques which truly abandoned these principles
would look so different from current techniques that it would be inappropriate to call them an
extension of current techniques. But what can we say about how these new techniques would
look?

What is needed

Please recall that I am not saying that there are no techniques which can handle dynamic
general language. I am saying that current techniques are insufficient. So what would be
sufficient? First, the new techniques would allow for fundamental ambiguity. Fundamental
ambiguity goes beyond superficial ambiguity in that it entails both an indeterminate list of
possible senses for a word and an indeterminate relation between the senses and the real world.
Most people in NLP to whom I pose the question of whether they believe in a universal set of
concepts determined by the structure of the universe will respond that they do not believe in
any such thing. Yet their techniques are based on this assumption. Again, within a domain, we
can act as if there were no fundamental ambiguity so long as we have a group of people who
have come to a shared understanding of the concepts of the domain. This shared understanding
comes about through human experts interacting in a mixture of general language and
specialised terms. General language provides the metalanguage for arriving at a common
understanding. But this approach falls apart when applied to general language, because there is
no metalanguage in which to discuss general language. Yorick Wilks has pointed out this
problem when he asked how one can know whether everyone in a co-operative effort has the
same understanding of the primitive concepts of an interlingua. This leads to the philosophical
problem of the given. How do we obtain the atomistic concepts that are used to build up more
complex concepts? What gives them to us? Chomsky would say that they are genetically
hardwired. Philosophers would say that if they are not hardwired and we do not have them as
children then we cannot get them through direct experience since concepts are required to
interpret our experience.

A satisfactory solution must overcome the problem of the given. Chomsky's solution is
unsatisfactory since it does not allow for fundamental ambiguity. One criterion that a
satisfactory solution must pass is the test of dynamic metaphor. Current NLP techniques can
easily handle frozen metaphor. We simply put a fixed expression in the dictionary. Although
even there we run up against resolving ambiguities such as the English request to go jump in
the lake or the French request to go cook oneself an egg, which may be literal requests to
perform a specific task or idiomatic requests to just leave and not come back, depending on the
situation. Dynamic metaphor is much more challenging than frozen metaphor. Dynamic
metaphor is created for the purposes of one text or even one sentence. Understanding dynamic
metaphor involves taking into account the entire situation and those aspects of general
knowledge that are relevant to the situation. It is ultra context-sensitive and thus contrasts with
the objectivist processing which assumes that the meaning of a sentence can be built up
without taking into consideration the context at all. Some dynamic metaphor is so clever or
poignant that it is frozen and preserved for future use. The prevailing wisdom is that metaphor
is a secondary aspect of language that can and should be ignored until other problems are
solved. Lakoff has shown that it is a pervasive aspect of language that needs to be solved up-
front. Certainly, for general language, we cannot afford to ignore it. An interesting aspect of
metaphor is that, although one cannot prepare in advance a list of all possible metaphorical
uses of a word and although once a dynamic metaphor is created one cannot predict how it
could be appropriately translated, every metaphorical usage is in retrospect motivated rather
than random.

Ian Kelly supplied me with an interesting sense history of the word 'treacle' over the past
two thousand years. At each change in sense, there was dynamic metaphor at play and each
change is motivated though some are surprising. The ancestor in Ancient Greek of the word
'treacle' was a wild animal. It then metonymically became the bite of a wild animal. This sense
then broadened to become a general injury and later shifted to the medicine used to treat such
an injury. Later still it narrowed to the substance put into a medicine in order to make it more
palatable and finally, in British English, to one such substance, molasses. Each step is logical
and motivated for a human, but it would be asking too much of a machine based on objectivist
assumptions to figure out the new meaning at any stage of the transition from wild animal to
molasses. Some NLP projects have worked on understanding dynamic metaphor. They should
not be expected to achieve human levels of performance unless they truly abandon their
objectivist assumptions. But at least it should be possible to measure their performance in such
tasks as translating texts containing dynamic metaphor.

Is there anything else that would be needed in a viable approach for handling dynamic general
language? Yes, it would be important to avoid falling into radical relativism when allowing for
fundamental ambiguity. Radical relativism, typified by the deconstructionist movement in
literary theory, recognises the problem of the given and solves it by saying that nothing at all is given.
Concepts are not genetic, neither are they built into the structure of the universe.
Everything is relative. The problem with this approach is that it does not explain how we can
communicate. How do we know that our concepts have anything to do with the concepts in
the head of the person we are talking to? A series of distinguished philosophers, including
Heidegger and, in his later work, Wittgenstein, have struggled with this problem. They have
concluded that our concepts are grounded in our social interactions. This is a promising
direction. Note that it implies that general-language linguistics is a branch of social rather than
individual psychology.

Often it is said that a computer that could translate anything would have to understand what
it is translating. But how do you tell if a computer understands? John R. Searle proposed a
puzzle ('Minds, Brains, and Programs', in The Behavioural and Brain Sciences, Vol. 3, (c)
1980, Cambridge University Press) in which it is assumed that techniques are somehow
developed which allow a person sitting in a box who speaks only English to answer questions
about a story by mechanically following a set of rules. The catch is that the story, questions,
and the answers are all in Chinese and the person is English monolingual. Within a domain, say
the domain of Chinese weather bulletins, this could probably be done if someone who speaks
only English could follow rules similar to those used by the Meteo system to translate weather
bulletins between English and French. Of course, it may take quite a while for the person in the
box to produce an answer, but let us ignore that problem. The question is whether the ability to
mechanically produce acceptable answers would constitute a demonstration that the person
understands Chinese. Most people would say the answer is obviously no, while strong-AI
people would say the answer is obviously yes.

Searle is on the side of those who think the answer is obviously no. He points out that if he were
the person following the mechanical rules, he would get out of the box without knowing any
Chinese. He would still know English and understand questions posed to him in English, but he
would not understand Chinese. He points out that some people have suggested that an adding
machine UNDERSTANDS arithmetic and that a door that opens automatically when someone
approaches it and breaks a beam of light UNDERSTANDS the instructions of the photocell. He points
out that this sense of 'understand' is not at all the same as the sense in which we note a person
understands Chinese. Searle then goes through several types of replies he has received to his
argument from strong-AI types. One type of reply is that perhaps a person that blindly follows
the rules sitting closed up in a box does not understand Chinese, but if the rules were
programmed into a small computer that was put into a robot, then the robot, thanks to its
ability to move about and see things, would understand. Searle replies that this implies that
understanding is solely a matter of formal symbol manipulation, which is one of the tenets of
objectivism. Searle counters the symbol-manipulation theory by noting that an essential
element of understanding is conscious intentionality. Most people would accept this. The
problem is how to detect whether a machine intends to do something or merely follows a series
of instructions. Strong-AI proponents must logically accept a form of mind-brain dualism,
namely that the mind, including its intentionality, can be successfully implemented in a digital
computer or in a human brain or, presumably, in 'Cartesian mental substance' if we ever run
across any of that stuff, whatever it is. Searle takes delight in pointing out, however, that the
AI literature contains 'frequent fulminations against "dualism".' Searle rejects this form of
dualism and expresses his belief that intentionality is a biological phenomenon. If this is so, he
points out, we should no more expect a computer program to have intentionality than a
computer software simulation of photosynthesis to produce sugar. The problem is that unless
we can somehow detect intentionality and prove that it is a biological phenomenon, we have a
stand-off between Searle and Minsky. They may agree that a computer needs understanding and
that understanding entails intentionality, but that leaves unanswered the question of whether a
computer can have intentionality. In line with my previous stance of attempting to identify
specifically what would be needed for a computer to handle general language rather than just saying
that it would need an undetectable spark, we should perhaps look for indirect ways to detect
understanding and intentionality.

Marvin Minsky, in the same article where he pontificates about artificial brains, says something
with which I agree, namely, that one thing which separates current machines from humans is the
flexibility of the human mind. When a computer program encounters a situation for which it has not
been explicitly programmed, it stops or produces meaningless results. When humans encounter a
new situation, they are able to try various solutions until something works. This applies to Searle's
Chinese Box puzzle. Flexibility is a detectable aspect of understanding and intentionality. Even
Meteo has occasional problems with a sentence, usually due to a typographical error or noise on the
transmission lines. A human reviser handles these situations because they cannot all be systematized
and therefore require the flexibility of the human mind. The human blindly following the instructions
of Meteo would exhibit no more flexibility or robustness than a computer. Therefore, neither a
computer, nor a human following instructions mechanically, truly understands. We have now made
an additional requirement of a machine that might handle natural language. It must exhibit flexibility
in handling new situations. This flexibility would probably be related to the ability to handle dynamic
metaphor. It seems that ways of testing flexibility could be devised.

Joseph Weizenbaum is well-known for having written a computer program called Eliza that
simulates a psychoanalyst. When it was first installed on a computer at a university, some people
would TALK to it for hours on end through a computer terminal, exposing their darkest secrets
and actually believing that it was a human psychoanalyst or at least that it really UNDERSTOOD
them. Weizenbaum was appalled. ^{(note 8)} He knew that the computer program didn't understand a thing
they were saying. It simply looked for key words and put together minor variations on stock replies.
For example, if a person said: 'My parents are divorced', Eliza would reply something like: 'Tell me
more about your family', using a table that lists 'parent' as a 'family' word. How was Eliza so
successful in fooling intelligent people? First, it was dealing with a domain, the domain of the
detached psychoanalyst gathering data. Eliza never said anything substantive, even mundane things.
It could not even answer a question like 'How many days are there in a week?' It just asked
questions to keep the person talking, and who doesn't like to talk when SOMEONE will listen? Eliza
clearly fails the flexibility test of being able to handle a new situation.

It is instructive at this point to look at one other person who has written about mechanistic
approaches to language. Roy Harris in his book The Language Machine (1987) traces the history of
the idea that human language can be put into a machine, going back to Gulliver's Travels in which
there is a section about a machine which randomly produces sequences of words. Young men are
employed for the purpose of sifting through the random sequences for ones that have meaning and
putting the sequences together into books. This satire on a wrong way to create literature is
surprisingly not too different from the deadly serious way that a Chomsky-style grammar randomly
produces sentences in isolation to supposedly generate a human language, except that it is semantic
rules that eliminate the millions of sequences that do not make sense instead of a room full of
humans, a process euphemistically called 'overgeneration' and 'selection'. Along the way we find
Saussure who posited a language machine in the brain in order to distinguish linguistics from
language teaching. For him, the language machine was automatic and so no one had control. Thus
there was no need to teach the inner workings of the language machine to humans and no danger of
language teachers taking over part of linguistics. What is missing from Saussure is any mention of
bilingual humans or social class differences in dialect. They were erased by the idealized langue.
Saussure spoke out against prescriptivism, but, ironically, it was during Saussure's lifetime that the
idea of a standardized national language arose, a triumph of prescriptivism, with theoretical support
from Saussure's idealization of language. A national language is a creation which gives a false idea
of uniformity and contributes to the view of language as a machine. Then Chomsky completed the
project by making language into a machine that functions completely without human intervention.
For Harris, the view of language as a machine has contributed to the exclusion of a moral dimension
from language and a devaluing of a search for solid truth and knowledge, resulting in radical
relativism. Another bizarre consequence of the language machine view is that communication is
only an incidental aspect of language instead of the core aspect.

There is a contradiction between the model of language as a machine that is independent of social
interaction and the deepest yearnings of the authors of these models. Chomsky, in a documentary
on his life and work, stated that, although he has sought a connection between his linguistic theory
and his political activism, which centers on manipulation of public opinion by the press, he has found
none. ^{(note 9)} Shouldn't that lack of connection be worrisome? And Minsky, in his article about artificial
brains, makes the rash claim that 'No popular ethical system yet, be it humanist or religion-based,
has shown itself able to face the challenges that already confront us.' He is clearly concerned about
the meaning of life for himself and others. He even ends his article with a sermon-like plea: 'Our job
is to see that all this work shall not end up in meaningless waste.' It seems that a good place to start
would be to place social interaction at the core of language and give some of the long-established
ethical systems a chance to work instead of undercutting them.

Now we can put together the previous points of flexibility and social interaction to avoid both the
problem of the given and radical relativism. We need a flexible grounding of language that allows
for social interaction at the core of language. This leads to the work of Levinas on the questions of
interiority and totalization. Levinas has shown that to be an interiority, that is, to have selfhood and
agency, which is also an essential part of consciousness and understanding, one must acknowledge
the existence of other interiorities that cannot be totalized. Put in more familiar words, to be a living,
thinking person, one must acknowledge the existence of other thinking, living persons who are
peers and agents themselves and whose actions and motives cannot be perfectly controlled or
predicted. Even an attempt to control other people is an implicit acknowledgment of their agency
which you wish to destroy. Totalization involves bringing something into your world and gaining
complete control over it. Totalization in an NLP system that interacts with people in any way would
involve making a model of the person with which the system is interacting and incorporating that
model into the algorithm of the system, so that the computer actually interacts with the model,
which is part of itself, not with the person. ^{(note 10)} An algorithm is a finite set of instructions such that
each decision is binary (i.e. yes or no) and the process terminates in a finite number of steps. All
computer programs (except those stuck in an INFINITE LOOP are algorithms. Once we accept
that, although totalization of the physical world is desirable and largely possible, totalization of other
people is neither possible nor ethical, then we can draw the startling conclusion that an approach to
dealing with natural language that truly allows for social interaction could not be a totalized system
and therefore could not be algorithmic! For entirely different reasons, a prominent physicist Roger
Penrose (1989) has suggested that the brain may operate non-algorithmically on the basis of faster-
than-light processes of quantum mechanics.

This brings us back to the image at the beginning of the paper. The oval represented a knot-hole in
a tree house. The tree house could stand for a domain-specific approach. Useful work can be and is
accomplished in machine translation with a domain-oriented approach based on the assumptions of
objectivism. However, before computers will have a chance of performing as well as humans on
dynamic general language, they will at least have to avoid the assumptions of objectivism, allow for
fundamental ambiguity, handle dynamic metaphor, become much more flexible, and become an
agent, recognizing other people as agents (which involves being based on a non-algorithmic
approach). The final step of becoming an agent and seeing others as agents that regard it as an
agent, thus permitting social interaction, is suggested by the step of looking along the beam of light
(in the tree house story told earlier) instead of at it. Until you do it, it is impossible to know what
the result will be. Once you do it, a whole new world opens up.

Implications

The implications of this philosophical discussion are simple. Machine translation is headed in
the right direction. Domain-specific approaches using controlled language should be continued
and the controlled languages should be made to conform to all the assumptions of objectivism
so far as possible. Dialogue-based machine translation can guide the user into writing in a
controlled language. Low-quality indicative translation for information only is unarguable since
many find it useful. But further work on fully-automatic high-quality machine translation of
unrestricted text is a waste of time and money unless the issues in this paper are carefully
addressed. If we ever reach a breakthrough in natural language processing which allows for the
handling of dynamic general language, it will not be based on any extension of current
techniques in machine translation. The electric light bulb did not result from research and
development on the candle (personal communication from Roger Harris). Fully-automatic
high-quality machine translation of unrestricted text will be a truly surprising, unpredictable
breakthrough and therefore is not expected in the foreseeable future, even though it may come
at any time.

We should not complain about the heavy requirements I have imposed on an approach that
could handle general language at human levels of performance. In 1984, many of us reviewed
the vision of the world presented by George Orwell in his novel Nineteen Eighty-Four and
were thankful that things were not as bad as he had predicted, at least outside the Soviet Bloc
in the Free World. I had occasional contact with people on the other side of the Iron Curtain
and heard horror stories of oppression heaped upon those who dared think on their own in a
way that opposed the government then in power. In Orwell's world, the Party had invented
Newspeak, a deliberately restricted language in which it was impossible to think thoughts that
were not approved by the Party. Now ten years later, we have seen the Iron Curtain fall. If all
language suddenly could be treated like domain-specific language, then a new and far worse
Iron Curtain would, in Orwellian fashion, forever keep us from thinking truly new thoughts and
we would become machines trapped in the prison of objectivism.

(The ideas in this paper are more fully developed in Alan Melby's book, The Possibility of
Language, published by John Benjamins in the Translation Library Series. Full bibliographical
references may also be found in this work.)

Notes 1.. Terry Winograd (1987) provides an additional example of the fact that meaning is not always neatly divided up into a literal base meaning and figurative extensions. Suppose one asks the question ' Is there any water in the refrigerator?' In the context of a typical American family this would be a question about whether there is a pitcher in the family refrigerator containing enough cold water (above zero degrees Celsius but probably below ten degrees) to pour into a glass and have a good drink. However, a scientist asking another scientist this same question may be asking whether there is any substance in the laboratory refrigerator containing some H2O that might interfere with an experiment using microwaves. Which is the literal meaning? If one tries to list all the possible meanings in all conceivable contexts, this is an admission that meaning is indeed dependent on context. If one argues that the literal meaning is the one that is most likely in a normal context, then this is also an admission that meaning is dependent on context in this case, the context we have called the Utterly Boring World. There really is no meaning that is independent of all context. [BACK] 2.. There are, of course, even variations in the legal system between states in the United States and between England, Wales, and Scotland in Great Britain. Further complications arise when considering US territories and extra-British members of the United Kingdom, such as Northern Ireland, the Isle of Mann, and the Channel Islands. [BACK] 3.. John Hutchins commented on the shift from general machine translation systems to domain-specific systems at the 1994 Cranfield conference. At that same conference, Peter Wheeler, who in the past ten years has gone from working at the European Commission with Systran, to working for Logos (a machine translation developer), to being an independent consultant, confirmed the accuracy of the remarks made by Hutchins. [BACK] 4.. At the first conference of the Association for Machine Translation in the Americas, a member society of the International Association for Machine Translation, held in Columbia, Maryland, October 6-8, 1994, a panel discussion treated the topic of the future of machine translation. Several panel members expressed their belief that current systems would gradually be extended to handle general language. [BACK] 5.. At the 1994 Cranfield conference, I took a straw poll during a debate on the limits of machine translation in which professionals from all over the world were participants. About ten percent of the participants indicated that they take the strong-AI position. [BACK] 6.. An outrageous observation at this point would be that there seems to be something about people whose names end in 'sky' (pronounced 'skee') that leads them off the deep end. [BACK] 7.. An outrageous observation at this point would be that there seems to be something about people whose names end in 'sky' (pronounced 'skee') that leads them off the deep end. [BACK] 8.. I learned how chagrined Weizenbaum when I heard him give a lecture on the topic in the 1970s. [BACK] 9.. This statement was made in the film Manufacturing Consent, a documentary on the life of Chomsky which has been shown on university campuses and art film theaters around the country. [BACK] 10.. This suggests another way to detect understanding. Ask someone to make friends with the computer program. Have them ask the computer for advice and try to determine whether the computer program really cares about the person or is just 'going through the motions.' This would test both flexibility and interiority. [BACK]

References

Lakoff, George (1987) Women, Life, and Dangerous Things: What Categories Reveal about
the Mind, University of Chicago Press.

Winograd, Terry (1987) Computers and Cognition: a New Foundation for Design, Wokingham: Addison-Wesley.

Click here for

EXIT

top   end

next item   previous item

 (end)