Semantic Networks in Language Understanding

As we noted earlier, semantic networks have been used to represent the meanings not only of words but of whole sentences. We saw earlier, for example, that the semantic net formalism can be used to distinguish between the two 'deep' interpretations of the predicate "is yellow" in sentences like "Canaries are yellow" and "Butter is yellow". Early networks developed for the representation of sentence meaning include those of Robert Simmons (1973) and his co-researchers and successors; Nick Cercone (1975, 1980); and the LNR group (Rumelhart, Lindsay & Norman, 1972; Rumelhart & Norman, 1975). All derive their ideas ultimately from the linguist Charles Fillmore (1968), who devised a case-based grammar that captured similarities and differences in sentence meaning missed by phrase-structure grammars. Consider, for example, the sentence "John paints nudes". Although there is one preferred interpretation, at least two are in fact possible: that John produces artistic representations of naked human bodies, and that John applies paint to naked human bodies (see figure 6.7.). A case grammar is able to show up this distinction by associating with the main verb of a sentence a set of conceptually required roles, the whole constituting a `case frame'. The verb `cook', for example, has two conceptually required roles: some agent who is doing the cooking and some object which is being cooked. With the correct assignment of cases, a case grammar is therefore able to capture the semantic relation between the following three sentences:

     Maria is cooking the chicken
     The chicken is cooking
     Maria is cooking

While the second and third sentences have the same `surface' structure, the case frame for `cook' shows up their difference at a `deep' conceptual level. Further, since only sentence elements filling the same case-role can be conjoined, case grammar also accounts for the ill-formedness of the sentence below:

     *Maria and the chicken are cooking

Cases other than Agent and Object that are commonly used in case-based network representations of sentence meaning are: Instrument, Result, Source, Path, Goal, and Experiencer.

Figure 7.

Weisberg (1969), in experiments on sentence processing similar in intent to those of Quillian & Collins on taxonomic trees, provides psychological evidence that appears to support a case-based model of language comprehension.

Quillian's hierarchical classification of world knowledge had no place in the original networks of Simmons' and company, which, like Fillmore's case grammar, were designed to capture the meanings of sentences by extending from a node representing the main verb a set of links to nodes representing the cases associated with the verb. The networks were held to capture the way in which we store meanings in memory, and by which we recognize when two or more sentences are, despite their different structures, semantically equivalent. Thus, to take a simple example, the meaning of the sentence "John gave Mary an apple" might be represented by a net such as:

Figure 8.

where the top node names the type of action, the central node indicates that this is a specific occurrence of the action type, and the remaining nodes are labeled by the names of the participants in the action. The links between the nodes indicate the relationship of the participants to the action: on this particular occasion of giving (that is, give_37), it is John who does the giving, the apple which is given, and Mary who receives the apple. Furthermore, the arrows on the links indicate how the meaning should be read: that for example, the apple is the object of the act of giving and not that giving is the object of apple. Consequently, the same network also captures the meanings of other sentences such as:

   John gave an apple to Mary
   Mary was given an apple by John
   It was John who gave an apple to Mary
   It was an apple that John gave Mary
     ... and so on.

thereby showing that, with respect to the propositions that they express, the sentences are paraphrases one of the other.

Because, in any act of giving, the underlying case relationships remain the same -- there is always someone who gives, something given, and someone to whom it is given -- we might also see the network above as the filling in, for the sentences listed above, of a more abstract schema. That same schema also provides an interpretation for numerous other sentences: "The boy gave his mother flowers", "Charlie gave his son a blank cheque", "Othello gave Desdemona a handkerchief", ... and so on. We might then say that the meaning of each of these sentences is `give', together with specific values filling the case slots, as below:

instance_of: giving_events
agent: some human, e.g., John, the boy, Othello
recipient: some human, e.g., Mary, his mother, Desdemona
object: some object, e.g., apple, flowers, handkerchief

Among the developments from this form of the semantic network which deserve mention, though we shall not discuss them further here, are Shapiro's distinction between different kinds of relations between nodes, Woods' (1975) proposals for dealing with quantification (as in "Every student failed an exam") and Hendrix's `partitioned semantic networks', and Schank's conceptual dependency representation which translated natural language sentences into their underlying conceptual forms, expressed as conceptual primitives.