COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	AN EPISTEMOLOGICAL APPROACH TO NATURAL LANGUAGE
C00011 ENDMK
C⊗;
AN EPISTEMOLOGICAL APPROACH TO NATURAL LANGUAGE


	Human language can only express facts about the world
that humans can obtain.  Indeed what we can express is only a
subset of what we can know.  All natural language is constrained
by these human limitations.  Moreover, it seems that artificial
intelligences will be subject to some of the same limitations and
not to others.  The object of this paper is to study the restrictions
that our ability to acquire knowledge puts on human language.
A further object is to devise an "artificial natural language" based
on first order logic that can be used by a robot to express
internally and communicate what it can find out about the world.
The study may be regarded as a part of "applied epistemology".

	We begin with some observations about these limitations.

	1. The world is composed in a large measure out of 3-dimensional
solid objects which often have important internal structures and
which are usually opaque.  We can see the parts of their outsides
that face us and aren't occluded by other objects or parts of the
same object.

	1. We cannot give accurate estimates of the dimensions of
objects without measuring them.

	1. We don't have access to our raw visual data.  Instead we
see objects and shapes, and the neural processes that generate
this information are not directly observable.  We may compare this
with a hypothetical robot that gets visual information by a television
camera that puts an array of pixels into its memory.
  The robot may also be provided with object
recognizers that put their output in memory.  However, the robot
may also look at the pixels themselves if this is important.

	Since the robot can be constructed otherwise, it is an
accident of human evolution that we have access only to highly
processed visual data rather than a consequence of the intrinsic
nature of observation, i.e. it is a matter of neurophysiology rather
than philosophy.

	1. Much that we can see is inexpressible in language without
special conventions or equipment.  Suppose I want to tell you that
I see green at a certain place in my visual field, leaving aside
the fact that this is not an elementary fact about a pixel, because
it depends on contrasts.  I can do it if I can tell you that the
clown has a green nose, i.e. if the place is identifiable as a
namable part of an object.  I can also do it if we establish a
co-ordinate system with the aid of instruments such as a telescope
with a reticle.  I can also give the location of the place relative
to other namable objects provide there are such conveniently located.
Moreover, people whose work requires describing  the location
of objects to colleagues develop this ability more than the rest of
us.

	1. The ability to describe faces well enough so that another
person can recognize them is not common but can be acquired by some
people with practice.

	1. We aren't very good at estimating and comparing intervals
of time but can do better with instruments and with practice.

	1. The terms we use are not well defined.  Even the most
common terms become controversial when examined closely.  We believe
that this means that some vagueness is essential for 

	1. At first a baby does not identify its actions with the
corresponding actions of other people, so it has no need of an
internal name for itself.  Thus its own  going  has a different
name from someone else's  going.  When it comes to make the parallelism
which permits conscious imitation and also reasoning from itself to
others, it needs a private internal name for itself.  When it begins
to talk it needs a public name for itself and uses whatever its mother
teaches it, e.g. "baby" or "Johnny".  Later it learns to use the word
"I" which is needed for communicating with people who don't know its
name and for the imitation of others who use "I".  Making emotional
attachments to the word "I" occurs still later.

	We can ask how much of this would be needed by a robot.
Since we are building an adult robot, we don't need the early stages.
It also doesn't matter whether the robot uses "I" or its name
in internal sentences.  For external use it will need both, since
these are appropriate in different conversational situations.  Even
if its communication were only with similarly designed robots, we would
need both "I" and a name unless we assume that the name is visible
and attached in all interactions.  Of course, its use of "I" could
be kept minimal by having it say first "I am Smith" and then making
all subsequent references to itself by name.  It could even say
"Smith here".  Likewise it would be convenient to design in other
pronouns.

	The pronouns of sex are convenient but not necessary, and
many languages don't distinguish "he" and "she".  The Japanese
experience is illuminating.  They didn't distinguish until they
began translating Western literature into Japanese.  In Western
literature many sentences contain both "he" and "she" which is
convenient when interactions between people of opposite sex are
being described.  In order to translate these sentences without
circumlocution, the Japanese invented a word for "she".  Note that
circumlocutions are required when the interactions are between
people of the same sex, so the convenience of "she" is only available
part of the time.

	1. Consider the difference between languages with and
without articles.  Languages without articles distinguish by
context most of the time and use words with meanings like "this"
and "a certain" when the distinction is important and the context
doesn't provide the information.