LAWS@SRI-AI.ARPA (11/12/84)
From: AIList Moderator Kenneth Laws <AIList-REQUEST@SRI-AI>
AIList Digest Sunday, 21 Oct 1984 Volume 2 : Issue 143
Today's Topics:
Programming Languages - Buzzwords,
AI Tools - LISP Machine Benchmarks,
Linguistics - Language Evolution & Sastric Sanskrit,
Seminar - Transformational Grammar and AI,
PhD Oral: Theory-Driven Data Interpretation
----------------------------------------------------------------------
Date: 19 October 1984 22:52-EDT
From: Herb Lin <LIN @ MIT-MC>
Subject: buzzwords for different language types
Could someone out there please tell me the usual catch phrases for
distinguishing between languages such as C, Pascal, Ada on one hand
and languages such as LISP on the other?
Is it "structured" vs "unstructured"? List vs ??
Thanks.
------------------------------
Date: Fri 19 Oct 84 13:08:44-PDT
From: WYLAND@SRI-KL.ARPA
Subject: LISP machine benchmarks
A thought for the day on the AI computer benchmark controversy.
We need a single, simple measure for machine quality in order to
decide which machine to buy. It must be simple and general
because these are typically intended to be used as general
purpose AI research machines where we cannot closely define and
confine the application.
We already have one single, simple measure called price. If
there is no *simple* alternative number based on performance,
others (i.e. those funding the effort) will use price as the only
available measure, and we will have to continually struggle
against it using secondary arguments and personal opinion.
It should be possible to create a simple benchmark measure. It
will - of necessity - be highly abstracted, necessarily crude.
This has been done for conventional computer systems: the acronym
MIPs is now fairly common, for good or ill. Yes, there are
additional measures, but they are used in addition to simple ones
like MIPs.
We need good, extensive benchmarks for these machines: they will
point out the performance bugs that are unique to particular
designs. After we do the benchmarks, however, we need to boil it
down to some simple number we can use for general purpose
comparason to place in opposition to price.
------------------------------
Date: 19 Oct 84 10:32 PDT
From: Schoppers.pa@XEROX.ARPA
Subject: The Future of the English Auxiliary
In response to Ken Kahn's question on language evolution, my own theory
is that the invasion of a language by foreign cultures, or vice versa,
has a lot to do with how simple a language becomes: cross-cultural
speakers tend to use only as much as absolutely necessary for them to
consider themselves understood. The English spoken in some communities,
eg "Where they goin'?" (missing an auxiliary), "Why he be leavin'?"
(levelling the auxiliary), "He ain't goin' nowhere" (ignoring double
negatives), etc may well be indicative of our future grammar. On the
other hand, "Hey yous" for plural "you" (in Australia), and "y'all"
(here), are pointing towards disambiguation. Well, there does have to be
a limit to the simplification, lest we "new-speak double-plus ungood".
Then again, "ain't" can mean any one of "am not", "aren't", "isn't",
"haven't", "hasn't" --- effectively replacing both the primary English
auxiliaries (to be, to have) in all their conjugations! United States
"English", being the lingo of the melting pot, will probably change
faster than most.
Marcel Schoppers
Schoppers@XEROX
------------------------------
Date: Fri 19 Oct 84 15:23:26-MDT
From: Stan Shebs <SHEBS@UTAH-20.ARPA>
Subject: Cases & Evolution of Natural Language
Has anybody at all researched the origins of language? Not an expert
on the subject, but I do know that the languages of aboriginal tribes
are extraordinarily complicated, as languages go. But they probably
don't give us much clue to what the earliest of languages were like.
If you believe that the earliest of languages arose along with human
intelligence, then you can suppose that the most primitive languages
had a separate "word" for each concept to be expressed. Such concepts
might include what would correspond to entire sentences in a modern
language. Thus the most primitive languages would be completely
non-orthogonal. When intelligence developed to a point where the
necessary vocabulary was just too complex to handle the wide range
of expressible concepts, then perhaps some individuals would start
grouping primitive sounds together in different ways (the famous
chimpanzee and gorilla language experiments suggest that other primates
already have this ability), resulting in the birth of syntactic
rules. Obvious question: can all known languages be derived
as some combination of arbitrarily bizarre syntactic/semantic rules?
(I would guess so, based on results for mathematical languages)
Word cases can then be explained as one of the last concepts to
be factored out of words. In the most ancient Indo-European languages,
for instance, prepositions are relatively infrequent, although
the notions of subject, object, verb, and so forth have already
been separated into separate words. Perhaps in the future, singular
and plural numbers will be separated out also (anyone for "dog es"
instead of "dogs"?).
stan shebs
------------------------------
Date: 19 Oct 1984 15:17-PDT (Friday)
From: Rick Briggs <briggs@RIACS.ARPA>
Subject: Sastric Sanskrit
Firstly, the language is NOT artificial. There is a LITERATURE
which is written in this language. It is different from toy artificial
languages like Fitch's in that for three thousand years scientists
communicated and wrote texts in this language. There are thus
two aspects which are interesting and relevent; one is that research
such as I have been describing was carried out in its peculiar context,
the other is that a natural language can function as an unambiguous,
inference-generating language without sacrificing simplicity or
stylistic beauty.
The advantage of case is that (assuming it is a good case system)
you have a closed set with which a correspondance can be made with a
closed set of semantic cases, whereas prepositions can be combined in
a multitude of ways and classifying prepositions is not easy.
Secondly, the fact that prepositions are not attached to the word
allows a possibility for ambiguity "a boat on the river near the
tree" could be "a boat on the (river near the tree)" or "a boat (on the
river) near the tree". Attaching affixes directly to words allows you
(potentially) to express such a sentence without ambiguity. The Sastric
approach is to allow one to express a sentence as a series of "facts",
each agreeing with "activity". Prepositions would not allow this.
If one hears "John was killed", some questions come to mind: who did
it, how, why. These are actually the semantic cases agent, instrument,
and semantic ablative (apaadaanakaaraka). Instead of "on" and "near"
one would say "there is a proximity, having as its substratum an
instance of boatness... etc." in Sastric Sanskrit. The real question
is "How good a case system is it?". Mapping syntactic case to semantic
is much easier than mapping prepositions since a direct correspondance
is found automatically if you have a good case system, whereas
prepositions do not lend themselves to easy classification.
Again, Sanskrit is NOT long-winded, it is the english
translation which is, since their vocabulary and methodology was more
exact than that of English.
"Caitra cooks rice in a pot" is not represented ambiguously.
Since it is not specified whether the rice is boiled, steamed, or fried
the correct representation should include the fact that the means of
softening the rice is unspecified, and the language does have the
ability to mark slots as unspecified (anabhihite). Actually, cooking is
broken down even further (if-needed) and since rice is cooked by boiling
in India, that fact would be explicitly stated. The question is how deep
a level of detail is desired, Sanskrit maintains: as far as is necessary but
"The notion 'action' cannot be applied to the solitary point reached by
extreme subdivision", i.e. only to the point of semantic primitives.
Sentences with ambiguity like "the man lives on the Nile" in Sastric
is made up of the denotative meaning (the man actually lives on the
river) and the implied meaning (the man lives on the bank of the Nile).
The latter is the default meaning unless it is actually specified
otherwise. There is a very complex theory of implication in the
literature, but sentences with implied meanings are discouraged because:
"when purport (taatparya) is present, any word may signify any meaning",
thus the Sastric system where implied meanings are made explicit.
I do not agree that languages need to tolerate ambiguity,
in fact that is my main point. One can take a sentence like
"Daddy ball" and express it as an imperative of "there is a
desire of the speaker for an unspecified activity involving the ball
and Daddy." By specifying what exactly is known and what is unknown,
one can represent a vague mental notion as precisely as is possible.
But do we really need to allow such utterances? Would something
humanistic be lost if children simply were more explicit? Children
in this culture are encouraged to talk this way by adults engaging
in "baby talk". All this points to the fact that the language you
speak has a tremendous influence on the your mental make-up. If
a language more specific than english was spoken, our thoughts would
be more clear and ambiguity would not be needed.
I conclude with another example:
Classical Sanskrit--> raama: araNye baaNena baalinam jaghaana (Rama
killed Baalin in the forest with an arrow) --->
raamakartRkaa araNyaadhikaraNikaa baaNakaraNikaa praaNaviyogaanukuulaa
parokSHaatiitakaalikii baalinkarmakaa bhaavanaa (There is an activity
relating to the past beyond the speaker's ken, which is favourable to
the separation of life, which has the agency of Rama, which has the
forest as locus, Baalin as object, and which has the arrow as the
implement.
Note that each word represents a semantic case with its instantiation,
(eg., raama-kartRkaa having as agent Rama), with the verb "kill"
(jaghaana) being represented as an activity which is favourable
(anukuulaa) to the separation (viyoga) of praana (life). Thus the
sentence is a list of assertions with no possibility of ambiguity.
Notice that Sanskrit expresses the notion in 42 syllables (7 words)
and English takes 75 syllables (43 words). This ratio is fairly
indicative of the general case.
Rick Briggs
------------------------------
Date: 19 Oct 1984 15:41 EDT (Fri)
From: "Daniel S. Weld" <WELD%MIT-OZ@MIT-MC.ARPA>
Subject: Seminar - Transformational Grammar and AI
[Forwarded from the MIT bboard by SASW@MIT-MC.]
Transformational Grammar and Artificial Intelligence:
A View from the Bridge
Robert Berwick
It has frequently been suggested that modern linguistic theory is
irreconcilably at odds with a ``computational'' view of human
linguistic abilities. In part this is so because grammars were
thought to consist of large numbers of explicit rules. This talk
reviews recent developments in linguistic theory showing that, in
fact, current models of grammar are quite compatible with a range of
AI-based computational models. These newer theories avoid the use of
explicit phrase structure rules and fit quite well with such
lexically-based models as ``word expert'' parsing.
Wednesday October 24 4:00pm 8th floor playroom
------------------------------
Date: 19 Oct 84 15:35 PDT
From: Dietterich.pa@XEROX.ARPA
Reply-to: DIETTERICH@SUMEX-AIM.ARPA
Subject: PHD Oral: Theory-Driven Data Interpretation
[Forwarded from the Stanford bboard by Laws@SRI-AI.]
PHD ORAL: TOM DIETTERICH
DEPARTMENT OF COMPUTER SCIENCE
2:30PM OCTOBER 25
SKILLING AUDITORIUM
CONSTRAINT PROPAGATION TECHNIQUES FOR
THEORY-DRIVEN DATA INTERPRETATION
This talk defines the task of THEORY-DRIVEN DATA INTERPRETATION (TDDI)
and investigates the adequacy of constraint propagation techniques for
performing it. Data interpretation is the process of applying a given
theory T (possibly a partial theory) to interpret observed facts F and
infer a set of initial conditions C such that from C and T one can infer
F. Most existing data interpretation programs do not employ an explicit
theory T, but rather use some algorithm that embodies T. Theory-driven
data interpretation involves performing data interpretation by working
from an explicit theory. The method of local propagation of constraints
is investigated as a possible technique for implementing TDDI. A model
task--forming theories of the file system commands of the UNIX operating
system--is chosen for an empirical test of constraint propagation
techniques. In the UNIX task, the "theories" take the form of programs,
and theory-driven data interpretation involves "reverse execution" of
these programs. To test the applicability of constraint propagation
techniques, a system named EG has been constructed for the "reverse
execution" of computer programs. The UNIX task was analyzed to develop
an evaluation suite of data interpretation problems, and these problems
have been processed by EG. The results of this empircal evaluation
demonstrate that constraint propagation techniques are adequate for the
UNIX task, but only if the representation for theories is augmented to
include invariant facts about the programs. In general, constraint
propagation is adequate for TDDI only if the theories satisfy certain
conditions: local invertibility, lack of constraint loops, and tractable
inference over propagated values.
------------------------------
End of AIList Digest
********************