WHY THERE MUST BE A SEMANTIC REPRESENTATION (OVER AND ABOVE 
ANY COG~TIVE ~ETWORK) 
Th. R. Hofmann 
Toyama University, Japan 
A s~nantic representation (or a semantic network; Mel'- 
~uk, Seuren, Hofmann, Sgall, ...) cannot be dispensed with in 
a model of language comprehension which incorporates a re- 
presentation of knowledge, commonly called a oo~itive net- 
work (e.g. Qu£11ian, Lamb, Shank, Hays, Jackendoff, ...). 
We shall demonstrate this in several ways, claiming to re- 
solve the contention between those WhO claim they are necess- 
ary (the 1st camp above) and those who would ols/m that they 
can be dispensed with, the 2nd oemp, including also Hontague, 
who uses them for"convenience only". 
First, and most intuitively, is the observation that 
.hi can and commonly does understand a description of some- 
thing (e.g. a scientific theory, or a political tract) which 
one knows or believes to be at variance with the facts. This 
suggests that in-taken information is kept apart from general 
knowledge, until such time as one decides to accept it aa 
true. Cognitive networks can, however, be augmented to account 
for this by annotating arcs by their (episteeLic) source, 
though this is a bit unrealistic for most human use of langu- 
age - we seldom know the source of some believed fact. 
A 2nd demonstration is to note that the norms~ person 
"knows" of many different worlds, where facts are at varian- 
ce with those of an other world. As a scientific example, 
Riemsnian and Eucledian geometries are such contrary worlds, 
- 134 - 
as are any pair of compet~ theories. For more ordinary 
examples, most English speakers know of at least 4 different 
worlds, conta£ning different objects, different possibilities, 
and so on: the world of @reek mytholo~ (conta4nLng unicorns, 
gods, etc.), the world of Sherlock Holmes (with a certain Dr 
Watson, a Baker St, ...), the world of James Bond, and the 
"real" world. As for the real world, one's interlocutor often 
has a different version of it, which must be known to under- 
stand his speech. 
Now, if one must incorporate in himself knowledEe of 4 
or so d£stinot worlds (and I would suggest it is closer to 40 
than 4), then we may say he has as me~y cognitive networks, 
of v~dus degrees of d ete£1 and completeness. These networks 
may be represented as conflated into one super cognitive netr 
work, with sufficient special marking of the type eu~ested 
above (extended to nodes, also), but such a oonflated network 
can always be decomposed into separate ones for the venous 
worlds, if it is adequately marked to model human behawlour. 
If then there are a number of distinct cognitive net- 
works needed for understanding ordinary human speech, there 
is no reason not to add 1 more for the conversation currently 
in progress. This is no more nor less than a semantic repres- 
entation, except as we shall observe below, it differs in 
structure, and function. 
A tighter, but longer and more difficult, demonstration 
is tO show that "inferencing", which is roughly equivalent to 
moving through a cognitive network, is NOT undertaken until 
the sem~utio operation of integration is attempted. This 
operation, which combines the semantic contributions of 
sentences together, depends heavily on a principle whereby 
each successive sentence is interpreted in the most redundant 
way possible, so that it "overlaps" as much as possible with 
the content of the previous sentences. 
- 135 - 
t~ 
With this principle, we can show that inferencing 
appears to be undertaken only when integration is blocked for 
lack of averlap of the expressed meanings of the component 
sentences. (Inferencing is of course required for argument, 
for determ, tnAng truth of an assertion, or for otherwise apply- 
ing the comprehension of a language act against the world.) 
These "expressed meanings of the discourse" are nothing more 
than their sen~ntic representation, ld~ich we have thus shown 
to b9 necessarily held distinct from background knowledge 
used in lnferenoing, represented in a cognitive network. 
We can conclude, then, that in some form or other, a 
representation of the semantic effects of the sentences is 
needed to account for how discourses are understood. Although 
this semantic representation me~ be conflated (in a computer, 
e.g.) with a co~j~itive network, it is logically distinct, 
and may be profitably so-treated. 
This sems~utic representation, we may observe, is dist- 
inct from the ordinary cognitive network in being extremely 
undetai~ea contaihing mostly syntagmatic rather than pars- 
dl~natlo relatiofiships, and very malleable, and it has a 
priva'ledged position in the interpretation of language acts. 
It is also used pri~ns~rily, at speech time, as a depository, 
with perhaps no inferenci~, while a co~litive network as 
generally understood receives perhaps, nothing at speech time, 
but ie used primarily for inferencing. Thus they appear to be 
quite distinct in contents, function and usage. 
136 - 
