Further notes about CardWorld
James A. Mason - 2010 June, 2012 August
http://www.yorku.ca/jmason/asdindex.htm
Why CardWorld is not a
"toy" example
CardWorld is an example from a sub-field of
computational linguistics which can be called experimental
linguistics or language simulation. It
differs from the related fields of theoretical linguistics,
psycholinguistics, and language engineering in that its
objective is to implement detailed working models of
cognitive mechanisms by which humans understand and use
natural languages.
- Its pragmatic domain, though simple, feels quite "natural" to
a
human user and can be extended in various directions. It
is a
minimal but reasonably self-contained pragmatic domain.
Playing
cards are all the same size, and although cards are almost
two-dimensional, they can cover one another, making for
interesting
non-permanent collections. Most people already know about
cards
and how to manipulate them. They know that cards are
two-sided
flat objects that can be moved around, turned over, shuffled,
and
stacked. The first implemented version of CardWorld,
CardWorld1 (like its refined implementation CardWorld1a), does
not
even require knowledge about playing cards, because it only
deals with
cards as simulated physical objects. The next implemented
version, CardWorld2, does require knowledge of standard features
of playing cards (suits, ranks, colors, face cards, spot cards,
and jokers) as they are used in English-speaking countries.
- Because the pragmatic domain is reasonably simple, it can be
modeled thoroughly, with few additional simplifying
assumptions.
That
permits the focus of the CardWorld project to be on modeling the
syntax, semantics, and pragmatics of English-language
interactions with
it or, more accurately, with a simulated card agent
playing the role of a person. Yet the pragmatic domain is
rich
enough to support many aspects of English syntax and semantics,
including many extensions beyond the aspects modeled in
CardWorld1 and CardWorld2.
What CardWorld1 and CardWorld2 illustrate about and provide
working models of aspects of natural-language understanding
- Understanding a natural language is not an all-or-nothing
ability. In the beginning it does not require recursively
nested
syntax, the ability to perform deductive logic, and other
abilities
that are required for the full fluency of a well-educated
adult.
We know that, of course, from the language abilities of
children.
In most cases the CardAgent component of CardWorld1 understands
the
utterances that it accepts as well as, and in the same way as, a
person
would. To the extent that its understanding differs from
that of
most persons, it simply needs correction and refinement.
- The only requirements for communication by means of a language
are: (1) a shared pragmatic environment, (2) an ability to
exchange
utterance tokens, (3) vocabulary, parsing and interpretation
abilities
which have
enough in common between the interlocutors that the intended
pragmatic
effects of utterance actions can be accomplished.
CardWorld1
meets all those requirements.
- CardWorld1 models some specific subtleties about
direct-definite, deictic and anaphoric reference in English:
- The type of thing referred to, as well as the instance
referred
to, can be anaphoric. e.g. "Turn over this pile."
"Now
turn over that one." ("one" refers to "pile")
- Singular separate reference can be anaphoric. e.g.
"Turn
over each pile." "Then shuffle it and spread it
out." ("it"
refers to each pile separately.)
- Definite reference with the unique quantifier "the" can
refer
to the only instance, among a set of instances, to which the
action
specified by the verb can be applied. e.g. "shuffle the
pile"
when there is one multi-card pile and one or more separate
singleton
cards. (A pile consisting of a single card can't be
shuffled.)
- Definite reference with the unique quantifier "the" can
refer
to the unique non-degenerate instance among a set of
instances, all but
one of which are degenerate. e.g. "turn over the pile"
when there
is a multi-card pile as well as some singleton cards.
(This
differs from example 3, because singleton cards can be turned over.)
- Definite reference with the unique quantifier "the" can
refer
to an instance which is separate from one or more
collections of two or more instances of the same type.
e.g. "turn
over the
card" when there is one card alone by itself while the other
cards are
in multi-card piles.
- Reference with the unique quantifier "the" can be deictic if
it
would otherwise be ambiguous except that there is an instance
of the
type which has recently been singled out. e.g. move a
card, then
"turn over the card".
- CardWorld2 models some additional subtleties in reference that
involve shifts of attention from parts (cards) to wholes (piles)
and from wholes (piles) to parts (cards) with described
properties. Its software agent, CardAgent2, also models an
omniscient Card God (or, if you prefer, a Card
Superhero), which knows where all of the cards are on the
table regardless of whether they are face up, face down, or
stacked in piles, and which can manipulate them regardless of
where they are within piles. Later versions of CardWorld
will attempt to model also more-limited software agents, which
will be called Card Players, with more human-like
abilities.
CardWorld
home
page