Semantic Networks

Semantic Networks

This document concerns the management of the output of insight
generators, the software agents utilized in the insight generation
systems. The solution to managing these reports involves the automatic
creation of a repository for all materials generated by various
insight generators; this repository allows the user to navigate
through this continually growing space of marketing reports, gaining
new insights about the relationships between items of interest
and adding new insights in the process. The goal of the system
is to make all marketing information and insights generated by
the man/machine interaction available to the user, so that there
is a convergence towards a “conservation of information”.
To use a geometric metaphor, the goal is to make the user equidistant
from all information at all times, as illustrated below.

The output of insight generators like I Want is paper; every time
the system is run, a paper report is produced. If the system were
run for every retail account in every market, it would produce
hundreds of reports. If similar insight generators focused on
other insights (distribution, variety, coupons, shelf space, prices,
etc.), the collection of agents would produce thousands of reports.
If each agent were run for each brand item (each size, package
type, flavor, etc.), there could be many thousands of reports.
Finally, if all of the agents were run each month for all the
brands, then millions of reports could accumulate.

Further, these agents could be run for a firm’s competitors. The
system could be run backwards; that is to say, it
could be run from the perspective of the brand’s competitors in
a particular category. From this, the brand group could learn
where their brand is vulnerable to attack from competitors seeking
to take merchandising support away from it. Also, the sales force
could be informed where not to ask for more support, such
as in the instance where their brand is receiving far more feature
support than its volume share warrants. Such use of this system
can be called exposure analysis and it was further explored
in the Market Opportunity Inspector (MOI) prototype.

One approach to dealing with all of these insights is summarization,
which is explored in the Marketing Opportunity
Inspector (MOI) document. The MOI system assumes that an “I Want…”
system has been run for a brand in all accounts in all markets.
Each page of output corresponds to an exposure, an opportunity
or neither for the brand. MOI takes these and summarizes the number
and magnitude of the opportunities and exposures. The output of
the system is again a sheet of paper. MOI helps to generate insights
about a brand’s situation, by locating its strengths and its weaknesses.
The output of one insight generation system (I Want) became the
input of another insight generating system (MOI). In both systems
the output is a high-quality paper report. Insight generating
systems such as MOI are employed to summarize the lower level
information and pinpoint insights from this mass of textual output.
But, the summaries themselves also add to the ever-growing output
and must also be made available to users.

A second approach to the problem of insight management is to manage
the output itself as a sort of library of information known
about the brand. This library would be part of a system that manages
information and insights that are in the form of compound documents,
i.e. pictures and text, that goes beyond a strict hierarchical
arrangement. The user needs to be able to move in multiple directions
from any vantage point and create hierarchies as needed. The information
needs to be arranged in a structure that follows the intrinsic
relationships between the context of the reports and their contents.
The reports must be managed according to some “mental-model”
of the world of marketing.

One technology for capturing and reasoning with such mental models
is a semantic network … the topic of this document.

What is a Semantic Network?

Semantic networks are knowledge representation schemes involving
nodes and links (arcs or arrows) between nodes. The nodes represent
objects or concepts and the links represent relations between
nodes. The links are directed and labeled; thus, a semantic network
is a directed graph. In print, the nodes are usually represented
by circles or boxes and the links are drawn as arrows between
the circles as in Figure 1. This represents the simplest form
of a semantic network, a collection of undifferentiated objects
and arrows. The structure of the network defines its meaning.
The meanings are merely which node has a pointer to which other
node. The network defines a set of binary relations on a set of
nodes.

Figure 1

Pick up almost any technical book and look in the preface or introduction.
Invariably there is a chapter dependency diagram. It is a node-link
structure, a semantic network in which the nodes represent chapters
and the links represent the relationship of which chapters should
be read before which other chapters.

To move semantic nets from this abstract realm to something more
concrete, let us consider an example from the structure of marketing.
To begin simply, let us introduce two nodes and a link.

Figure 2

The node on the left labeled “Quad Cities” is linked
to the node on the right, labeled “Market”, and the
arrow is labeled “is-a”. Quad Cities is an example of
a market. The diagram, in other words represents the fact that
there is a binary relation between a market, Quad Cities, and
the concept of a market. Another node with the label “Los
Angeles” and a “is-a” link from this node to the
“Market” node could be added, again representing that
“Los Angeles” is a type of “Market”.

Figure 3

If a retailer node is added to Figure 2, the structure of the
network becomes apparent as shown in Figure 4. Markets generally
contain retailing entities. To add an example of a retailer, add
a node labeled “Chain56” and two links – one from the
retailer “Chain56” to “Quad Cities” labeled
“is-a-retailer-in” and one from the node “Chain56”
to the node “Retailer” labeled “is-a”. This
illustrates that Chain 56 is a retailer in the Quad Cities market.

Figure 4

It is now important to note a point or two of possible semantic
confusion. Notice that the nodes in this small network are not
all of the same “type”. The node labeled “Market”
represents the generic or meta or class concept of a market; it
represents the abstract concept of a market. It can be thought
of as possessing properties common to all markets. The node “Quad
Cities” represents an individual instance of the node “Market”.
The node “Quad Cities” represents a particular market.
The same is true of the relation between the node labeled “Retailer”
and the node labeled “Chain56”. The node “Retailer”
again represents the concept of a retailer that is common across
all particular retailers. One instance of such a retailer is the
node labeled “Chain56”. In order to distinguish between
these two types of nodes, the class nodes become boxes and the
instance nodes become ellipses, as in Figure 5.

Figure 5

Another class node, labeled “Item”, that represents
the abstraction of items in a category, can now be added. Along
with that, an instance of an item, labeled “87481”,
is added. Notice that there is a strong relationship between the
type or class nodes and the column headings or entities of a relational
database table. We will exploit this similarity later in this
paper. Thus, another “is-a” link and a new link, “item-carried-in”,
must be added to the node “87481” and the node “Chain56”
respectively. These new additions are shown in Figure 6. The information
now being represented is that Chain56 is a retailer in the Quad
Cities market and that Chain56 carries the item 87481.

Figure 6

As the nodes proliferate, the meanings of these links need to
be considered. It should become apparent that not all links are
alike. Some links express only relationships between nodes, and
are therefore “assertions” of the nature of the relationship
between two different nodes. For example, the link “item-carried-in”
in Figure 6, which illustrates the relationship that retailer
Chain56 carries the item 87481. The “is-a” links in
Figure 6 are “structural” links in that they convey
“type” information about the node. This information
is about the node itself and not about the relationship it has
to a different “type” of node. For instance, the node
labeled “87481” is an instantiation of the class node
labeled “Item”.

In Figure 7, more nodes and links are introduced to the original
network. There is now a “Brand” class node with an instance
node “Ivory”. The link “is-brand” conveys
the information that the item 87481 is the Ivory brand. There
are now also class nodes labeled “Manufacturer”, “Category”,
and “Category Attributes”. The Category Attributes class
is linked to three other class nodes labeled “Size”,
“Color” and “Segment”. These represent particular
attributes of a particular category; in this instance, the liquid
light duty detergent category, of which Ivory is a member. The
“is-a” links between the class node Category Attributes
and the class nodes Size, Color, and Segment represent a relationship
of class to subclass and, hence, “structural” links.
Here again are links that do not denote a relationship between
different types of instance nodes, but give information about
a class node itself. The class node Color is a type of Category
Attribute.

Figure 7

Our network in Figure 7 now has a representation for information
about the item node 87481. For instance, it is a form of Ivory
which is manufactured by Procter & Gamble; it is the 22 ounce
size, white in color and competes in the Mildness market segment
of the liquid light-duty detergent category. This is one item
in one chain in one market. The database used in the Marketing
Information Center prototype has 100 items in five chains in one
market. Each item can be one of seventeen brands made by twelve
manufacturers which comes in seven sizes and eight colors and
can compete in one of five segments. This database is a pared-down
version of a scanner database for the Quad Cities market which
has many more retailers and a good many more items. The network
in Figure 7 becomes very complex with a 100-fold increase in the
amount of information.

None of the networks have shown any structural links among class
nodes, except for Figure 7 which shows only a subclass relationship
between Category Attributes nodes and various class node attributes.
Figure 8 shows possible structural relationships between class
nodes.

Figure 8

In this figure the instance nodes have been left out in order
to show more clearly possible relationships between classes. Remember
class nodes represent larger, more general concepts and just as
general concepts can have more refined sub-concepts, the particular
types of category attributes, such as Size, are represented as
a sub-class of the broader concept of Category Attributes as shown
previously. Notice that the class Category Attributes is a kind
of abstract class that probably would never have an instance node
tied directly to it. It can only have relationships to other class
nodes. So, general concepts are represented such as the concept
that there are Manufacturers who create things termed Brands which
are suppied to things called Retailers. Retailers are in an abstraction
called a Market and carry instances of Items. All of this is obvious
from the diagram. What is not so obvious is that the nodes themselves
can contain more than meets the eye. The Retailer node is a short-hand
notation for a bundle of concepts that make up a real-live retailer,
such as the fact that retailers have a headquarters and stores
and control shelf space, price and display. The links, such as
the one labeled “supplies” between the Manufacturer
node and the Retailer node, are really more like a co-axial link
than a simple arc. This particular link represents a bundle of
various relationships between Manufacturer and the abstraction,
Retailer. This detail is shown in Figure 9.

Figure 9

Another import characteristic of the node-link representation
is the implicit “inverse” of all relationships represented
by the directional arrows. If there is an arrow going from one
node to another, this also implies the reverse – that there is
an arrow from the second node to the first. In Figure 10, there
are the nodes labeled “P & G” and “Ivory”
with the link labeled “makes”. The direction of the
relationship is that “P & G makes Ivory”. Further,
some linguistic terminology for our binary relationships could
be used: “P & G” is the subject and “Ivory”
is the object, and “makes” is the verb or action
or link between them. This will be discussed in greater
detail later.

Figure 10

This “P & G makes Ivory” relation implies the inverse
relationship that “Ivory is-made-by P & G”,
as shown in Figure 11.

Figure 11

The representational or expressive power of semantic networks
has been discussed thus far. As with any kind of knowledge representation
scheme, a way of inferring knowledge that is not directly represented
by the scheme is needed. The ability to work with incomplete knowledge
sets a knowledge representation apart from a database. To give
an example of what can be gleaned from the semantic network in
Figure 7 that is not directly represented, consider Figure 12.
It is an extraction of Figure 7 containing only three nodes and
two links.

Figure 12

The information explicitly represented is that the item numbered
87481 is the Ivory brand and that Procter & Gamble makes Ivory.
The inverse relationship of the item 87481 to the brand Ivory,
i.e. that Ivory is-item-number 87481 is shown in Figure
13.

Figure 13

By tracing the path from the node P & G to the node Ivory
via the arrow labeled “makes” and then from the node
Ivory to the node 87481 via the arrow labeled “is-item-number”,
we can infer that Procter & Gamble manufactures the item 87481
by inferring a link labeled “makes-item” between the
node P & G and the node 87481, as shown in Figure 14. This
may seem obvious, but remember this small amount of new information
need not be explicitly represented in the original network.

Figure 14

Described mathematically, composing arrows occurs by placing
them end-to-tail. This composition creates a new arrow.
In Figure 14, a triad of nodes is formed by arrows said to “commute”.
It is not possible to compose every pair of arrows, only those
whose destinations and sources correspond. The destination of
the first must be the source of the second. By composing arrows,
new relationships between nodes can be found and described. This
process is sometimes called “chasing arrows” and the
terminology introduced stems from a branch of mathematics called
Category Theory.

Figure 15 shows the results of more “arrow chasing”.
Additional relationships are derived, such as Ivory is a Brand
carried by Chain 56, Procter & Gamble makes a product that
competes in the Mildness segment,and Ivory is a Brand competing
in the Mildness segment. Notice that layers of relationships between
nodes can be built.

Figure 15

This discussion has introduced the concept of a semantic network
consisting of nodes and links. The nodes represent concepts and
the links represent relationships between these concepts. A distinction
was made between instance nodes and class nodes: the former represents
general notions of the latter of which there may be many types.
The concept of links which extend from the instance node level
to the class node level was given along with an introduction of
the notion of abstract classes. The reversibility of the arrows
and the method of inferring new relationships between nodes from
existing ones was also given. Several figures illustrated these
concepts using an example semantic network built from a scanner
database header file of the liquid light-duty detergent category.

Limitations of Semantic Networks

This chapter should not end without some discussion of the limitations
of semantic networks, and a comparison of their traditional use
versus their use in the management of marketing insights.

Semantic networks as a representation of knowledge have been in
use in artificial intelligence (AI) research in a number of different
areas. Some of the first uses of the nodes-and-links formulation
were in the work of Quillian and Winston, where the networks acted
as models of associative memory. Quillian’s work centers on how
natural language is understood and how the meanings of words can
be captured in a machine. Winston’s work concentrates on machine
learning and specifically on structural descriptions of an environment.
Winston’s work describes pedestals and arches formed from more
elementary pieces such as wedges and blocks; these make up the
famous “blocks world” that has been utilized by many
research efforts in semantic networks.

The other major area in which the use of semantic networks is
prevalent, is in models based on linguistics. These stem in part
from the work of Chomsky. This latter work is concerned with the
explicit representation of grammatical structures of language.
It is opposed to other systems that tried to model, in some machine-implementable
fashion, the way human memory works. Another approach combining
aspects of both the previously mentioned areas of study was taken
by Schank in his conceptual dependency approach. He attempted
to break down the surface features of language into a network
of deeper, more primitive concepts of meaning that were universal
and language independent.

Such creations and uses of semantic networks have led to any number
of epistemological problems. Numerous researchers have attempted
to address these problems. Barr and Feigenbaum state that:

In semantic network representations, there is no formal
semantics, no agreed-upon notion of what a given representational
structure means, as there is in logic, for instance.

Of course, the success of logic in this respect is debatable,
but semantic networks do tend to rely upon the procedures that
manipulate them.

For example, the system is limited by the user’s understanding
of the meanings of the links in a semantic network. As pointed
out previously in the example of the network of marketing insights,
links between nodes are not all alike in function or form. Hence
we need to differentiate between links that assert some relationship
and links that are structural in nature). A paper by Brachman
on the subtleties of the “is-a” link revealed even more
distinctions in the uses of this link. According to Brachman “is-a”
links can be divided into two groups, depending upon the nodes
involved. One use of the “is-a” link resembles our distinction
between an instance node and a class node. The link represents
the relationship of instance nodes to abstract, generic qualities
shared by many instance nodes. Brachman’s other use of a link
is between two instance nodes. These two major divisions can then
be further broken down into finer uses of the link.

If problems and sublime uses characterize links, then nodes are
not much better. The seeming simplicity of a node that represents
a single concept or object in the world is actually fraught with
complications. The question “what is a node?” is on
par with “what is a market?” As in the discussion of
links above, we need to distinguish between nodes that represent
some set of objects and nodes that represent classes of qualities
shared by these objects. There can be properties for instances
of a class that all members of a class share as well as properties
of the class itself. A property of a market is that it is made
up of retailers. For example, a member of the class Market, such
as Quad Cities, contains specific retailers, such as Chain 56.
Chain 56 is an instance of the class Retailer. The problems of
epistomology and the semantics of semantic network representations
are discussed further in Brachman.

It should be noted that the sample network discussed at length
in this book shares the advantages and disadvantages of any semantic
network. We take “naive semantic networks” aka “naive
set theory”. We do not worry overly about the epistemologico-semantic
problems associated with the use of the representation. Some of
the difficulties are in fact side-stepped since the final goal
for this network is not quite the same as the goals of other researchers.
One fundamental difference between the use of the network representation
in this book as opposed to elsewhere, is that we are not trying
to represent natural language or associative memory independent
of users. In fact we rely heavily on the user’s intuitive understanding
of the world. Semantic networks are used in this project as structures
of knowledge that encourage the user to interact with them.

Having marketing knowledge visible to the marketing professional
is one of the major advantages of AI technologies and of tools
such as spreadsheets. This topic is pursued in more depth in Chapter
18, while the next chapter describes a computer-based tool for
implementing semantic networks.