More roles
Patrick Cassidy <micra@tigger.jvnc.net>
From: Patrick Cassidy <micra@tigger.jvnc.net>
Message-id: <199509120336.AA22098@tigger.jvnc.net>
Subject: More roles
To: cg@cs.umn.edu, srkb@cs.umbc.edu
Date: Mon, 11 Sep 1995 23:36:19 -0400 (EDT)
X-Mailer: ELM [version 2.4 PL24]
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Content-Length: 11302
Sender: owner-srkb@cs.umbc.edu
Precedence: bulk
In the continuing discussion about how to represent roles, and
just what they are, most of the discussion has been at a relatively
abstract level, with the goal of finding a formal specification for
the "role" idea. One problem I have had in evaluating the ideas
presented is that I am not certain just what set of concepts the
discussants consider to be roles. In fact, I myself can't always
say, for specific concepts, whether I think they are or are not roles.
My main concern is in representing knowledge for language
understanding (as contrasted with reasoning with concepts already
converted into well-defined symbolic structures), and it may not
be surprising that I find the notion of role to be closely associated
with patterns of language use. In this thread of discussion, there
have been only a few examples of roles mentioned, and it might be useful
to examine a few more examples to clarify just what a role is.
In my own efforts to decide what to do with roles while trying to
classify the concepts in Roget, my observations suggest that
what I intuitively think of as a "role" constitutes a heterogeneous
collection of concepts which actually don't have single definable
thing in common. What Pat Hayes apparently believes, that the
notion of "role" should not have any special status beyond that
of other semantic relations, is a also tempting conclusion
from what I have noticed in the Roget arrangement. When
trying to specify the semantic relations between all those words in
Roget clusters, it is apparent that the semantic relations ones finds
are salient (associated with the concept being "defined") to
widely varying degrees. Those words which (I think) most of
us would consider as roles just happen to have semantic relations
which are stronger than the average association. But I can't find any
one property in common with all "roles". I will ilustrate this
problem with a few examples.
The prototypical "roles" of "father", "mother", and other
kinship roles seem to be those most commonly thought of as roles.
These may perhaps be the words most commonly cited as roles
because they usually occur in linguistic expressions adjacent or
very close to the "other object" to which they have the "role relation".
So, we usually say "my mother" or John's mother", but seldom just
"mother" without also mentioning who she is mother of. (There
are, of course, exceptions - "A group of mothers demonstrated in front of
..."). So by repeated association in ordinary discourse, we probably
internalize the notion that this word has a very special relation to
another word. And, when we say "mother of" we have clear notion that
the following word will be of type "child". The same tends to be true
of the names of officers of organizations. "President" is frequently
followed immediately by "of X" where X is the name of an organization.
Even when "The president" is used without immediately mentioning the
organization, it is usually mentioned nearby, or the context otherwise
makes clear what organization is intended.
But already in these two examples an important difference is
apparent. A person can never change a "kin" role, whereas people
in a "job" role can change their relation to that and other organizations.
Now, when we look at other "job roles" the strength of the relation
starts to fall off. We seldom say "plumber of" or "electrician of". It
conjures up no special relation. But it's still a job, which I
would like to imagine also is a sort of role, no? A plumber is
someone who installs plumbing, a mother is someone who bore a
child -- the plumber is creator of the plumbing system, the mother is
creator of the child. Are these roles deeply different? But they feel
different, I suspect, because they are used differently in linguistic
phrases, not because their conceptual structures have strong qualitative
differences. In this respect, I tend to agree with Pat Hayes that
"The idea of a 'role' is essentially a linguistic one".
The strongest associations of "job" words are with
the actions performed, not with other objects -- although typical
objects involved in their jobs may have strong associations
(the word "plumber" comes from Latin "plumbum", for the lead pipes
they used to use), but the linguistic co-occurrences are not as frequent
or as close as those of mother and child.
Take "student", which should be another "role". We can say
"student of", but the next word might be a teacher, or it might
be a subject. We can say "X's student", where X is a teacher,
but X would never be a subject of study -- for that we say "X student".
This role behaves different linguistically from the kin and officer
roles. Also, words for people who do certain actions may have something
in common with nominalized verbs, where "the destruction of" selects
for the patient of the verb "destroy", and so does "destroyer of".
But the patient relation is not usually thought of as a role-relation.
I suspect that this is because the patients of most verbs are not
strongly typed, and any given patient does not co-occur with its verb
as frequently as does the role-related word with a more typical
role word.
In contrast to things people do, words for people with certain
physical attributes do not conjure up a notion of role. When we
speak of a "giant", nothing other than the specifically indicated size
is saliently associated with the word.
Another problem: when someone "plays a role" does that mean that
s/he "has" that role? If someone "plays the fool" is that person
a fool? Is "fool" a role or a description of someone's essential
character? If one plays the "Devil's advocate" is that a role?
I am not certain that the "role" of "Play a role" is the same
as the roles "mother" or "president". A president doesn't have
to play a role, he *has* that role. A person "Playing a role" suggests
acting, or imitation. If an object "plays a role" it suggests
that the role is not its usual function. A brick may be used
"as" or "in the role of" a doorstop.
Parts of objects are another type of concept which has (for me)
an intuitive feeling of a "role". Both "X's part" and "part of X"
are likely to refer to the same (whole) object X. And parts are
often replaceable (even sometimes in people). But the "whole"
object may not always be mentioned in the same dialog with the part,
although it frequently (usually?) is.
When we consider artifactual objects, it is apparent that almost
all of them are made for a purpose, and the purpose for which they
are intended creates an association strong enough to qualify (to
me) as a "role". But the linguistic expression of this "instrument"
role is quite different from kinship or job. In English we say
"cut with a knife", so the case (role) marker is "with", not "of".
In Russian, a special case is dedicated mostly to the instrumental
role-relation. The major difference here is that the word related
to an instrument is usually an action, not another "object".
But it still has the feeling of a role, in that the same task might
be performed by instruments with a variety of characteristics (though
perhaps with some "core" characteristics). At the Montreal
workshop the question was raised as to what is a "chair" or a
"nail". Does the function of these object play the most essential
part in their definition. Consider "fastener". Many objects may
play that role.
The tempting conclusion is that "roles" are, as Pat
Hayes seems to feel, just another type of semantic relation. But
these particular semantic relations may seem special to us because of
the frequent direct association of their semantically related words in
the same sentence as the "role" word. The strength of my own
feeling that some specific word specifies a "role" seems to me to be in
rough proportion to how frequently the word occurs in sentences associated
with a strongly typed semantically related word. Associations within
a discourse but outside of the same sentence create less of an
intuitive feeling that some concept specifies a role. Thus my
present suspicion is that "roleness" is merely a strength attribute
of a semantic relation, i.e. a measure of how frequently a word
and its semantically related word co-occur closely in a sentence.
If so, this is likely to be a continuously variable strength, with
no clear line of demarcation of roles from non-roles -- unless
phrasal patterns are used to distinguish them.
This leaves open the question of whether fillers for syntactic
verbal cases (such as instruments) can be considered roles, or whether
there is any need to distinguish relations between objects from
relations between actions and their cases. Fritz Lehmann suggests
that the role-relation holds between objects, and I presume he would
therfore exclude verbal case relations, as well as attributes
of objects. It is true that phrases like "the coronation of
the king" and "the king's coronation" have a different feel from
"the son of the king" or "the king's son". But I don't know how
to exploit this difference for my own chosen task, language
understanding.
In my own efforts, I have decided that there is one useful
thing that can be done with the concept of "role", which is to specify
those words "x" for which the "x of Y" and "Y's x" phrasal patterns
create a strong default for the type of the role-filler Y. Such words
are not designated as subtypes of "role" -- it was pointed out
by Jan Schmidt early in this thread of discussion that a father
isn't a role, he plays a role. These concept types are thus specified as
having a "roleness" attribute (a clumsy but safe terminology). This is
used for nouns designating objects (which are not nominalized actions or
attributes). I expect that knowing such things about certain words
will be useful in language understanding and generation; whether such
"roles" are qualitatively different from other semantic relations
on a conceptual (as contrasted with linguistic) level, and whether
they also correspond to others' notions of what a role should be,
I don't know.
On a related but more general level, the strength of each semantic
relation, as measured by the frequency of co-occurrence of concept types
in text, is also likely to be useful in language understanding tasks
such as word sense disambiguation and semantic restrictions on
syntactical combinations. Such association frequencies are probably best
extracted from large text corpora.
All of this leads back to a variant on the question which started this
thread of discussion -- how would one indicate linguistic attributes,
such as frequency of co-occurence between words or concept types,
in a conceptual graph? Should conceptual graphs be used for such
lexically-oriented purposes, or is that beyond their intended functions?
Can we only specify that a relation exists, or can we make some
"meta" assertions about a relation between concept types?
___________
Patrick Cassidy
MICRA, Inc. || (908) 561-3416
735 Belvidere Ave. || (908) 668-5252 (if no answer)
Plainfield, NJ 07062-2054 || (908) 668-5904 (fax)
internet: cassidy@micra.com