Report on ANSI X3H4 meeting
sowa@watson.ibm.com
Message-id: <9202030318.AA06745@quark.isi.edu>
Date: Sun, 2 Feb 92 22:15:00 EST
From: sowa@watson.ibm.com
To: kr-advisory@ISI.EDU, srkb@ISI.EDU, interlingua@ISI.EDU, cg@cs.umn.edu
Cc: mzemanko@note2.nsf.gov, jwinkler@nasamail.nasa.gov, skperez@mcimail.com,
jfulton@atc.boeing.com, mjones@atc.boeing.com
Subject: Report on ANSI X3H4 meeting
The ANSI X3H4 committee on IRDS (Information Resource Dictionary Systems)
met in Orlando, Florida, from January 27 to 31. On Thursday morning,
three people from the Knowledge Sharing Effort (Richard Fikes, Bob Neches,
and Bill Swartout) spoke to the group (primarily the X3H4.6 Task Group on
the IRDS Conceptual schema). They presented an overview of the KIF, KRSS,
and SRKB projects and discussed possible ways in which the Knowledge
Sharing Effort could collaborate with ANSI and ISO.
At lunch after the presentation, Fikes said that he believed there was
about a 90% overlap between the KIF and KRSS languages and hoped that
they could eventually be merged. I made the point that conceptual
graphs have features like indexicals and syntactic annotations that are
needed for natural language processing, but that the first-order subset
can be mapped to and from a sorted FOL in nearly a one-to-one fashion
("nearly" means that if you do the mapping CG -> Sorted FOL -> CG,
the result is logically equivalent to the original, but not always
identical). If KIF had types, the CG-KIF mapping would be just as
direct. But without types, the mapping loses the information about
which one-place predicates were intended to represent types and which
were intended to represent monadic relations.
Fikes replied that KIF does have a metalanguage capability for saying
which predicates were intended to represent types. But he also made the
point that the idea of adding types directly to the KIF syntax has been
considered and that he was sympathetic to it. A large part, if not the
majority of the AI community has been using typed languages for years,
and the addition of types to KIF would go a long way towards satisfying
their objections.
We also discussed the metalanguage capability of KIF, which uses "holds"
and "value" to apply relation variables and function variables to their
arguments; i.e. if $r is a relation and $f is a function, then
(holds $r $a $b $c)
asserts that the relation $r is true for the arguments $a, $b, and $c;
similarly,
(value $f $a $b $c)
is a term that denotes the value of the function $f applied to the
arguments $a, $b, and $c. With "holds" and "value", it is possible to
get the effect of a second-order logic within a first-order framework.
For similar reasons, conceptual graphs use an equivalent technique to
get the effect of second-order logic: It is not possible for the same
symbol to occur in the referent field of one concept and the type field
of another concept or relation. Instead, there are two operators tau
and rho that map names of individuals to labels of types and relations;
e.g. if "red" is the name of an individual of type COLOR and "on" is
the name of an individual of type RELATION, the following concepts are
well-formed:
[COLOR: red], [RELATION: on].
But to talk about a ball with attribute red or a cat on a mat, it is
necessary to apply tau and rho:
[BALL]->(ATTR)->[tau(red)], [CAT]->(rho(on))->[MAT].
Using upper case to distinguish type and relation labels from names of
individuals of type TYPE or RELATION, we could write tau(red) = RED
and rho(on) = ON. Those two graphs could then be written
[BALL]->(ATTR)->[RED], [CAT]->(ON)->[MAT].
In effect, rho corresponds to holds in KIF. Since conceptual graphs
do not distinguish functions syntactically, they must be treated as
relations that are functional in the last argument: value would be
treated as holds with an implicit argument for the result. If types
were added to KIF, another operator with the same effect as tau would
also be needed.
We also discussed questions concerning contexts and microtheories, as
in Guha's recent extensions to Cyc. We generally agreed that contexts
would be an important feature to consider, but that there are still
many unresolved issues concerning them. This is a hot topic for future
proposals and for research to evaluate various alternatives.
Fikes and Swartout left after lunch on Thursday, but Neches stayed in
town for further discussions on Friday morning. One issue we discussed
was the idea of standards. Bob said that he had been in the unfortunate
position of talking about standards in front of an audience that was
overwhelmingly hostile to the very idea. He would prefer not to be
placed in such a position again. He would be happy to work with ANSI
and ISO, but he wishes that they could change the S in their names to
something less controversial.
However, Neches also had a long talk with Gary Roket from Boeing,
who has been lobbying very hard for standards. Roket is concerned with
information exchange among different departments at Boeing and between
Boeing and their suppliers, subcontractors, and customers. He wants
well-defined interfaces that their programmers can begin using as soon
as possible, and he wants all software vendors to use exactly the same
interfaces so that Boeing isn't locked into anybody's proprietary
software. Boeing is just one large, but typical example. Roger Burkhart
>From Deere & Co. is spending 50% of his time on standards committees,
including ANSI X3H4, PDES (Product Data Exchange), and OMG (Object
Management Group). Deere is willing to spend that much money on Roger's
salary plus travel because they believe that standards are essential to
their business.
Standards are coming, and there is nothing that AI researchers can do
to stop them. But there is a great deal that they can do to help guide
the designers in the right direction. SQL is a prime example: It was
originally defined by people who were right down the hall from Ted Codd's
office, but Codd did not take an active part. As a result, SQL evolved
into a form that he was not happy with, and he has spent many years
criticizing its shortcomings. But now it's too late. With a few months
of active effort in the design stage, Codd could have accomplished much
more than he has with many years of critical reviews in the later stages.
Task Group X3H4.6 also made a presentation on the status of the IRDS
Conceptual Schema to the full committee on Wednesday afternoon, and we
spent most of the rest of the week working on a draft of the forthcoming
technical report. A complete first draft is due in March, and the final
report is due in July. If the report is approved by X3H4, we would then
continue developing it into a Draft Proposed Standard. Ideally, we would
like to see a common semantic basis for conceptual graphs, for KIF, and
for PDES/SUMM (the Semantic Unification Meta-Model that has been proposed
for PDES). If we are successful in reaching such a common basis, the
semantics would become the core of the standard, with conceptual graphs,
KIF, and SUMM as alternative syntactic representations. Other syntaxes
could also be used to express the same semantics, including a version of
the Cyc epistemological level or a cleaned-up, logic-based version of SQL.
The next meeting will be in Renesse, Holland, from March 9 to 13 for the
ISO Special Group on the Conceptual Schema. The US delegation to Renesse
will be headed by Maurice Smith from the X3T2 committee on communications
and will include members of both the communications and IRDS committees.
The purpose of the ISO group is to develop a conceptual schema that can
accommodate the needs of communications, IRDS, databases, and CASE tools.
Nobody in ANSI or ISO is lobbying on behalf of knowledge bases, but I
hope that if we can develop a good logic base for KIF, PDES, and IRDS,
we would be in a strong position to propose it as a conceptual schema
for ISO.
Those people who are still afraid that standards might cramp their
style have a right to be concerned. But we must realize that standards
are coming, and there are only two alternatives:
1. A clean, logic-based standard that could serve as a neutral interface
between systems based on any kind of reasoning techniques, including
neural networks, fuzzy logic, or some super nonmonotonic method.
2. Another half-vast standard like SQL that is full of ad hoc
excrescences and limitations.
A logic-based standard would not impose a limitation on anyone's
creativity in developing new kinds of reasoning styles. On the contrary,
it would enable software developers to provide a platform of basic
reasoning modules. Then creative researchers could build new kinds of
reasoning techniques on top of them that could combine the basic knowledge
in wild and wonderful ways. Without standards, there is so much dog work
required to do the basic stuff that it is very hard to do anything
really interesting on top of it.
John Sowa