| Thread Links | Date Links | ||||
|---|---|---|---|---|---|
| Thread Prev | Thread Next | Thread Index | Date Prev | Date Next | Date Index |
|
Hello Dr. Fuchs et al,
Thanks for joining into this
discussion, and for bringing
along an impressive list of others
who are in the field.
I've changed the subject title
slightly to start a new thread,
since this feels like a new
discussion, but I've posted
both of your earlier emails into this
one to keep the train
of thought intact.
I've also interspersed my comments
from other threads we've discussed on
this list from time
to time.
Norbert E. Fuchs wrote:
-------------------------------------------------------
To give this thread a constructive direction, I invite these discussants – and anybody else – to propose notations for knowledge representation, specifications, semantic web etc. that fulfil the following basic requirements that in one way or other John Sowa, Rolf Schwitter, myself and others working in the field of controlled languages have stated: • support of formal methods • processable by a computer, at best tractable • accessible even to people without a background in formal notations Norbert E. Fuchs Department of Informatics University of Zurich Switzerland From my own perspective, I used ROSIE
as a controlled
NL to develop knowledge bases in
intel assessment,
planning, and diagnosis. I
found it VERY difficult to work
with, and reverted to lisp as soon as
I could.
Yet I agree with the ultimate
functional objective of
controlled NLs, that is, to let the
subject matter experts
relate their own knowledge, learn
from other SMEs, and
pass on their knowledge in a machine
executable way.
That experience led me to write my
own controlled NL
soon thereafter, though I haven't
released it yet, and
I won't get a chance to work on it
intensively for a few
more years until I retire.
Your third objective, accessible to
people without formal
method training, is what I find most
important about them.
I found that the SMEs used language
in the same way
we all do - as highly compressed
descriptions of real
situations and actions,
interpretations and beliefs. They
are very good at doing logical
thinking, but they require
much better linguistic support than
I've seen in any of
the controlled nls I've seen so
far. In particular, they use
words and phrases as shorthand
descriptions of things
they've discussed with other SMEs
previously.
Specifically, they REQUIRE a great
deal of polysemy,
ambiguity, modal reasoning, and
search capabilities.
Of course, they don't describe their
thought processes
in those terms, but that is my
interpretation of their
requirements nevertheless.
I understand that those capabilities
can be added on to
a controlled nl, but I question the
practicality of doing so.
Instead, I believe that a controlled
nl should be limited to
the user interface, while more
effective algorithmic methods
should underly the controlled nl to
provide these capabilities.
So I compared controlled nls with
modern OO programming
languages, like my favorite one -
Delphi. I can define a
single word in many ways, using
polymorphisms, controlling
lexical scope, using overloadings,
and inheritance with
virtual methods and properties.
Still, this is not enough polysemy
for the SME-level user.
Since WordNet has been so widely
accepted, it seems to
me that the lexical relations which
WordNet contains could
also be applied to support
polysemy. The noun classes
represented in WordNet should be able
to handle some
degree of paraphrase, and the verb
relations from research by
Beth Levin, followed by many others,
should be incorporated
to help resolve the semantics of
otherwise ambiguous
syntactic forms.
Since corpus analysis has proven so fruitful in NLP
experiments, and since so
much of SME knowledge is
documented in reports, books,
papers, and other accessible
formats, a corpus analysis
capability is needed by any
effective controlled nl for practical
extraction of knowledge
in volume.
Since SQL databases are so widely
used today, and since
the exercise of applying knowledge to
large databases is
the first business level
justification for controlled nls, I believe
that a capability for using SQL
database views, procedures
and tables is essential for realistic
application of controlled
nls. That should be provided in
the underlying layers, not
in the controlled nl itself.
And I believe, especially after using
ROSIE, that it
is absolutely essential
that some domain objects should be
capable of being referenced
in contolled nls, but
implemented
in detailed programming languages,
such as Delphi or C++,
so that effective emulation of domain
objects can be hosted.
Some capabilities are often very
difficult to achieve in any verbal form,
but easy to achieve in functional,
executable form.
Although formal logic has all the
power of a turing machine,
it is often extremely dense as a
representation of otherwise
simple concepts. So a
controlled nl that stands alone is very
likely to be a failure in any
realistically sized application.
----------------------------------------
Norbert E.Fuchs wrote:
Hello I received a message that contains excerpts of an obviously long discussion on controlled natural languages, and I would like to comment on two statements made during this discussion. Since the thread is nested very deeply, it is not easy to identify who said what. Please accept my apologies if I wrongly associate a quotation to an author. (Probably) John Sowa wrote: As I said to Jean-Luc, controlled NL and true NL are totally different things that just happen to look the same for some narrowly delineated subset. Doug Skuce's ClearTalk, Norbert Fuchs's ACE, Rolf Schwitter's PENG, and my CLCE are all designed as more readable front-ends for knowledge engineers, *not* for casual users. The second sentence is not true for Attempto Controlled English (ACE). ACE has been intentionally designed to be used by domain specialist, e.g. by end users, that may or may not understand the underlying formal basis of ACE. As a case in point, ACE has been used for medical documentation by people with a background in (dental) medicine, but with no background in logic or computational linguistics. This does not imply that end users can simply start using ACE. ACE must be learned, concretely its small set of construction and interpretation rules must be learned. In our experience, this takes approximately 2 days. More time is needed to get fluent, though. (Probably) Rich Cooper wrote: question-answering systems, like simplistic agents in the appliances. Has anyone published work about using controlled nls in this kind of environment? Here I was discussing the use
of Wierzbicka's NSM primitives as
a sort of lowest level "life
force", with controlled NLs wrapped around
it. The idea I want to
explore is to consider Luc Steele's approach of
a grounded, distributed,
language learning capability in appliance agents
(e.g., car dashboards) that
could learn some degree of language
competence from other
appliances, and from users, as the environment
around it changes over
time.
So my question was, "has anyone published work about using controlled
nls
in agents that interpret a
grounded, distributed environment to learn from each
other, as
Luc Steele described in his experiments?"
We did. There is our work on the Attempto reasoner RACE. See for instance N. E. Fuchs, U. Schwertel, Reasoning in Attempto Controlled English, in: F. Bry, N. Henze and J. Maluszynski (eds.): Principles and Practice of Semantic Web Reasoning, International Workshop PPSWR 2003, Mumbai, India, December 2003. Lecture Notes in Computer Science 2901, Springer Verlag, 2003 Reasoning is the straightforward
logic part; I'm concerned mostly with
extending that through a learning
mechanism, like Steele's, and possibly
with a language learner, like van
Zaanen's alignment based learning methods.
In another project, a master student at the Technical University of Dresden used ACE as a control language for an agent expressed in FLUX. G. Dawelbait, ACE as a Communication Language for a FLUX Agent, Master Thesis, University of Dresden, 2003 Recently, we used ACE as query language for MIT's Process Handbook and just published A. Bernstein, E. Kaufmann, N. E. Fuchs, J. von Bonin, Talking to the Semantic Web – A Controlled English Query Interface for Ontologies, Proceedings 14th Workshop on Information Technology and Systems, December 11-12, 2004, Washington, D.C. This sounds relevant; do you have a
copy you can email to this list?
In each of these cases, the Attempto Parsing Engine APE translated an ACE text into a discourse representation structure DRS as interlingua. In a next step, the DRS was translated into the appropriate formal language for the target system. This highlights one advantage of ACE: one controlled language can replace a plethora of formal languages. For more details and demos check www.ifi.unizh.ch/attempto. Norbert E. Fuchs Department of Informatics University of Zurich Switzerland Thanks for joining the
discussion!
Rich
|