[ CogSci Summaries home | UP | email ]

Woods, W. (1975), What's in a Link: Foundations for Semantic Networks, in D.G. Bobrow & A. Collins (eds.), Representation and Understanding, Academic Press; reprinted in, Collins & Smith (eds.), Readings in Cognitive Science, section 2.2.

Author of the summary: Jim Davies, 1998,jim@jimdavies.org

Cite this paper for:

p36: This paper is concerned with the meanings of the network structures in a semantic network. It will argue that existing methods are inadequate and that such networks should represent propositions without committing to their truth value. An adequate notation will be introduced.

p38: The linguistic view:

The philosopher view:
Both of these views are caricatures.

p39: Truth specifications should be determined by a function that assigns truth value to propositions in particular possible worlds. This is argued in Woods (1967). This is procedural semantics.

p43: There is more to a semantics than making a semantic representation. There also needs to be a connection to the thing referred to.

p44: What distinguishes semantic networks from other semantic representations (such as predicate logic or deep structure) is the notion of a link which connects facts into a structure. Not only does it represent the knowledge, but it links associatively so that human-like retrieval is facilitated. That such a thing is possible is on faith, it has yet to be shown (see author note 1).

p45: The "logical adequacy" of a notation refers to its ability to precisely, formally and unambiguously represent any interpretation of a sentence a listener might have. Two more things are necessary too:

  1. translation from the sentence
  2. ability to make inferences
A canonical form is a basic form that a sentence in natural language can be reduced to. Two statements are equivalent if they can be transformed by a series of equivalence transformations (that preserve the sentence's meaning) to the same form. This process can be seen as a search. If the sentences have a canonical form though, then you just turn both sentences into their unique form once. This is a motivation to have canonical forms. However, certain formal systems cannot have canonical forms. Thus the expensive search is necessary.

Here's another reason: IF we want to know that e1 is implied by e2 Then we would check to see if the canon if e1 was already in the database. But this would mean that when we put in something we would have to put in all the implications. It is unlikely that it could work out this way.

Further, there is ambiguity in English. If Bill is Tom's uncle, then is Bill the brother of the mother or father? To solve this problem you can just represent uncle directly.

Partial canonicalization might be useful, though.

p49: The extension of a concept is the set of all such things in the world. The intension is its anstract meaning. For example, "morning star" and "evening star" have the same extension (venus) but not the same intension.

p52: Attributes and values are problematic too. Saying That Jon's height is 6 feet is okay, but what about representing that his height is greater than 5 feet, or that his height is greater than Sue's? Using intension we can solve this problem: Have one node for Jon's height and another for Sue's. Then connect them with a greater link.

p56: With three items you need to make further nodes. Like X is between Y and Z. You need a location node that holds the intensio "between Y and Z." This works for binary compositions, but with things like "Jon sold the book to Mary" it is problematic.

Fillmore's (1968) notion of a case representation is interesting. e.g.

  verb        sell
  agent       Jon
  recipient   Mary
  patient     book
this is setting up a particular instance of selling.

p59: The following is ambiguous:

  superclass    telephone
  modification  black
Is it an instance of a black phone or a statement that all phones are black?

p60: One bad way to deal with relative clauses is to treat them as though they were two propositions that share a node. e.g. "The dog that bit the man had rabies." The problem with this is that there is no way to tell which is the main clause. There is also nothing to tell us that these two propositions go together. How do we know both are true at the same time?

p62: Quillian and others have suggested that the relative clause is temporary, to be used to know how to place the referent of a statement when updating the system. But this ignores trying to capture the meaning of the sentence. Also, if you don't know about the dog, then you have to create a new node that has rabies. That was called the transient process account.

p67: Perhaps the strongest argument for intensional nodes comes from want statements. Jon wants a wrench. This does not presuppose the existence of a wrench. In conclusion, there need to be nodes that represent descriptions of entities as well as nodes that represent entities. It's still hard though-- representing the sentence "every boy needs a dog" is hard because the dog is variable (a different one for every kid) and the fact isn't positing the existence of any dogs anyway. So in total we need:

p71: Representing quantified expressions is usuallly done incorrectly. For example, to represent "every integer is greater than some integer" you might do this:

  verb  greater
  arg1  d2
  arg2  d3

  noun  integer
  mod   every

  noun  integer
  mod   some

But this cannot distinguish between "for every integer there exists some integer for which it is greater" and "there exists some integer that is greater than all integers."

p74 The author solved the problem by making the nodes quantifiers:

  type      quant
  quanttype every
  varianble x
  class     integer
  restriction T
  prop      s2

  type      quant
  quanttype some
  variable  y
  restriction T
  prop      s3

  type      proposition
  verb      greater
  arg1      x
  arg2      y
The cost of this is that the links are not as direct.

There are two other completely different ways to do it.

1. Skolem functions

convert all quantifiers to non-negative forms. Change all existential quantifiers into functions that create an instance. These functions take as arguments the universal quantifiers that came before it. So this:

for all x there exists a y such that for all z there exists a w such that P(x,y,z,w)

gets transformed into:

p(x, f(x), z, g(x,z))

The difficulty with this is that it is hard to represent negations.

The other way to do it is with lambda expressions.

Other things to worry about:

Summary author's notes:

  1. ACT-R and other architectures have since shown evidence that certain semantic networks can effectively model human information retrieval.

Back to the Cognitive Science Summaries homepage
Cognitive Science Summaries Webmaster:
JimDavies ( jim@jimdavies.org )