eventChemistry ™ .
Determining
Functional Load using SLIP
by,
Paul Prueitt, PhD
Founder, (1997) BCNGroup.org
President, OntologyStream
Inc.
Draft: December 27, 2001
The concept of functional load is addressed at some length
in John Lyons’ Book “Introduction to Theoretical Linguistics” (Cambridge
University Press, 1968). In Lyons’
book, the notion of functional load is treated as a cause of the distribution
of basic compositional elements related to spoken and written expression. This notion is a part of the tradition in
theoretical linguistics and follows the work of de Saussure (Course in General
Linguistics, Payot, 1916) and Z. Harris (Methods of Structural Linguistics,
Univ. Chicago Press, 1951).
In essence, the notion is that sounds that are easy to make
will be used in situations where ambiguity of expression has some penalty. So a basic investigation, on auditory and
acoustic phonetics, leads to an understanding of how language is used and
evolves. Auditory and acoustical
coherence and discordance is reflected in the structure and form of natural
language. The investigation leads to
partial knowledge when the target of investigation is a complex system.
My background is not so strong on theoretical linguistics
that I cannot proceed right away into a discussion of phonetics and grammar and
make comparisons between the internal structures of various language
groups. I will leave this to
others. In any case, this is not the
primary purpose of the OntologyStream Inc (OSI) Browsers. However, we suggest that these browsers can
be used as teaching tools and investigation tools related to distributional
analysis in general terms. In this way,
distributional analysis leads to the notion of event chemistry even if the
target of investigation is not linguistically focused. The event chemistry has to have:
1)
A theory of substructure (illustrated by the making and hearing
of sounds in Lyons’ perspective on functional load)
2)
Laws or rules of assembly, even if these laws are distributed
and not precise
3)
Expectancy
Natural language can be considered as a complex system that
is in fact stratified into organizational levels related to how language is
remembered, how the human body generates cognition and language expression, and
how natural language is reinforced within a social system. But natural language is only one example of
a complex system.
OSI Browsers are being developed to study the following
complex systems:
Clearly Natural Language Processing will provide some tools,
particularly with #3 and #4, where it is essential that concepts and themes
expressed in natural language is a significant part of the study.
But NLP work presents us with a dilemma. This dilemma is well stated by Lyons at the
end of his second Chapter:
“Two apparently contradictory principles has been
maintained in this section: first, that statistical considerations are
essential to an understanding of the operation and development of languages;
second, that it is in practice (and perhaps also in principle) impossible to
calculate precisely the information carried by linguistics units in actual
utterances. This apparent contradiction
is resolved by recognizing that linguistic theory, at the present time at
least, is not, and cannot, be concerned with the production and understanding
of utterances in their actual situations of use (except for a relatively small
class of language-utterances which can be handled directly in this way), but
with the structure of sentence considered in abstraction from the situations in
which actual utterances occur.” – page 98.
The link analysis made using the Warehouse
Browser provides a weak measure of functional load when the induced metric
is distributed and used to control an emergent computing process in the
scatter-gather in the SLIP Technology
Browser. A stronger measure of functional
load is possible only as domain experts begin to develop specific understanding
of the event atoms and the patterns of co-occurrence that is seen in
practice. This encoding can be
facilitated through knowledge management principles.
What one expects is the development of a type of periodic
table of elementary event atoms. The
development of this table is based on experimental results involving the
derived relationships between occurrences of atoms in event logs. In computer intrusion work, these atoms
might be IP addresses or port values.
In text understanding, the atoms may be co-occurrences of words in
paragraphs or other text units.
Figure 1:
Atoms from one of the SLIP categories
In Figure 1 we show
the event atoms for a category derived from a quick study of the functional
loads of the Aesop collection. As in Latent Semantic Indexing (LSI), we
focus on a relationship between the membership of an “internal token” and a
profile of the larger group. In this
case the token is a member of a set of referent
tokens (a very simple, and hand made, dictionary) and the larger group is
the collection of individual fables.
Table 1: the data set
for Figure 1
Looking at Figure 1 we see that atom 260
(one of the fables) has only two valances, and atoms 222
has three. By “valance” we mean here
that the Analytic
Conjecture has established an inference regarding how fables are related to
each other via the Dictionary of tokens.
This fact is reflected in Table 1.
Looking at concept linkage and functional load
In the simple exercise, to follow, we look at the concept
linkage between the elements of text in a collection. The concepts are weakly represented by a collection of nouns and
verbs that have been extracted from the fable collection. Functional load is to be identified through
what ever means we can. The first step
towards obtaining a validated theory on the functional load in the fable
collection is to build a first approximation using the co-occurrence between
individual fables and a unified list of nouns and verbs (called the
dictionary). A datawh.txt file was
produced for this purpose in the previous exercise.
A deeper study of functional load related to the fable
collection can be made. One could, for
example, parse the fables and identifying when a noun and a verb from the
Dictionary were both contained in the fable.
The events reported out to a new datawh.txt would then have the form
( noun, verb, fable
name )
Then the analytic conjecture could be done using nouns as
the “a” value and verbs as the “b” value.
But in this Exercise, we keep things very simple for illustration
purposes.
Table one is the Report generated by the Technology Browser
for category ‘R-D-level’ (residue at the D level).
a
b
Figure 2: The
Analytic Conjecture for tokens, and a SLIP Framework
In working with this exercise, we have two resources. First the collection of fables are posted one
URL at a time using the sting:
(www.) +
ontologystream.com/IRRTest/fables/BEAD(N).HTM
where “(N)” is replaced (manually) by the atom number. We have checked a few of these, but not
all.
The idea, with the fable collection, has been to prototype a
BeadGame Communities software
system based on work considered and made over a period of a decade by the
BCNGroup.org Foundation. However, this
long-term goal requires that our group make economic gains first. These gains will come from the application
of the NLP Browser to the analysis of patent information and stockholder
reports.
So we use the fable collection as a means to illustrate what
we have by way of technology and where there are still software development
issues that need to be solved.
The reader is now asked to download this exercise’s zip
file. The zip file for this exercise is TAI.zip.
(289 K zipped, including the three browsers and a data set.). The code is Visual Basic developed by
OntologyStream.com.
After opening the Warehouse Browser (SLIPWhse.1.2.0.exe),
one will see Figure 2a. Opening the
Technology Browser (SLIP.2.2.3.exe) will show the user an computer interface
that looks like Figure 2b.
The tree like structure, called the SLIP framework, seen in
Figure 2b is developed by taking the elements of the category A1 and randomly
scattering these elements (called SLIP atoms) to the circle. One can review the numerous exercises to review
how these SLIP atoms are developed. The
SLIP atoms are the “a” values from the analytic conjecture that where found (by
the Technology Browser) to have been paired with a “b” value. In this case, the SLIP atoms are names of
fables (labeled by the fable number).
So let us see how the SLIP atoms are created.
If you have not already done this, download TAI.zip
and unzip into an empty folder. You may
remove and delete the contents of the Data folder except the single file
datawh.txt.
Figure 3: TAI.zip
unzipped
Then click on SLIPWhse.1.2.0.exe . Enter the commands, “a = 1” and then “b = 0”. Enter the commands “pull”, followed by the
command “export”.
You can then look into the Data folder and see that several
new files exist, one of which is Paired.txt (used by the Technology Browser)
and Links (used by the Event Browser.)
Now click on SLIP.2.3.3.exe. The Technology Browser starts with an empty A1 node in the topic
graph window. Enter the commands
“import” and then “extract”. One can
type in “help” to find out more about these functions. Then click on the A1 node.
The development of the SLIP Framework can be automated. But
currently we insist on the user developing the Framework by visually finding
areas of interest by looking at the emergent clustering that occurs on the
circle. Enter the command “random’ and
then the command “cluster”.
The cluster command will produce 100,000 iterations of a
seek function. For this data, with this
analytic conjecture, this is too much iteration. Enter the command “random” and then “cluster 2”. Entering a return should add 2,000
iterations each time. By the time you
get to 12,000 iterations your gather process should look something like Figure
4b.
a b
Figure 4: Clustering the top node
We are looking for a small group of atoms. A small group helps we point out specific
features of the event chemistry related to that small group. We may take the middle out of the large
distribution and look into the complement.
To do this you may type in a bracket command “x, y -> B1”, where x
and y bracket the interior of the large distribution. Then click on the A1 node and type “residue” to put the
complement of B1 into a second category.
Now click on the residue category, labeled “R”, and type
random. Without clustering, take 90
degrees (any 90 degrees) and put the atoms into the category C1. To this by typing, for example, “45, 135
-> C1”. Click on the new
category. If you type “cluster” you
will most likely see the quick formation of several groups that do not move
together.
Choose the largest of these (hopefully you will have at
least two and less than five elements in this group.) If not, then go to the data folder and delete subfolders of A!
and type in load into the command line.
This will allow you to start over.
Once you have a category with two or three or four or five
elements in the category, then type in “key = 1”, click on the Report button
and type generate in the command line.
Figure 5: the
Report for a category of two atoms.
You will have different results than I. However, you have captured two – five atoms
that
1)
Are all connected by the co-occurrence of one or more token
and
2)
Have an average number of valances that connect to atoms
outside the category.
The two fables that I have identified here are (191) The
Jackdaw and the Doves, and (113) The
Master and His Dogs . The linkage is
via the token “food”.
Table 2: The valances
of 191 and 113
