REPERENCE RESOLUTION AND SEMANTIC COHERENCE 
Elisabeth Leinfellner, Ingeborg Steinacker and Herald Trost 
Department of Linguistics, University of Nebraska, Lincoln, 
Nebrasoa and Department of Medical Cybernetics, University 
of Vienna, Vienna, Austria 
The problem of how to define semantic coherence in a 
semantic net is a least twofold: 
(a) how should we structure the semantic net as to permit 
the combination of lexical meanings with the semantic 
structure of the vocabulary as such D 
(b) how do we go on from here when we want to explicate se- 
mantic coherence within clauses and texts? 
FOr ~hte specific problem it is important to show how 
- and how far - semantic coherence tallies with syntactic 
coherence and whether we need certain transformations in 
order to insure tallying. 
For our purpose, namely, to develop a conceptual natur- 
al language understanding system for German (VIE-LANG) we 
used Brachman°s (Brachman 1978) outline of a semantic net to 
construct a conceptual knowledge base. We included ideas from 
a "semanti.clzsd" valence theory, a "semanticIzed" dependency 
~smmar, and linguistic semantics in general. 
We will now give a short description of the semantic 
net as used in our project. There is, at first, a hierarchic- 
al "layer" of concepts, corresponding to a categorical the- 
saurus system. Each concept is structured via so-called 
roles through other concepts and it is here where the lexical 
meanings find their place. Roles either resemble deep cases 
- 171 - 
in a case grammar - so far, we have identified about 50 cases 
- or they are other vital "constituents'of the semantic con- 
cept to which they belong (e.g. properties). Whereas in the 
hlerarchioal "layer'the relations are those of hyponomy/hypero- 
moray, the relations which lead from concepts via roles to 
conceptual restrlotlons may be interpreted as relations of 
semantic compatibility (selectional restrictions), to be 
specified. 
There is no explicit dichotomy between syntax and se- 
mantics. Syntax is seen as a "coarse" semantic (pre-)ordering 
- with loopholes which have to be bridged by transformations. 
Each concept in the semantic net, including its roles and 
conceptual restrictions, can, therefore, be interpreted either 
as a semantic definition, or as a semantic sentence pattern. 
The latter approach accounts for the applicability of a °se- 
mantlcIzed" valence theory and a "semanticized'dependency 
grammar. 
How do we apply this system to the problem of semantic 
coherence in clauses and texts? The first practical step 
consisted in olasslfylng the forms of semantic coherence and 
integrating them into our system. Accordingly, semantic coher- 
ence is expressed via 
1. Proforms 
1.1. Profornm which can be identified morphologically and 
morphologlcally/syntactlcally 
1.1.1. Purely morphologically: lexemes which can be identif- 
ied as proforms on a purely morphological basis (per- 
sonal pronouns, certain adverbial proforms). 
1.1.2. Morphologically/syntactically: lexemes which can be 
identified as possible proforms on a morphological 
basis but we need syntactic evidence for the decision 
of whether they really are proforms: possessive pro- 
nouns (function as proforms only when isolated - non- 
-attributive), certain adverbs which may also funct- 
- 172 - 
ion as conjunctions, certain interrogative pronouns 
which have the same morphological shape as relative 
pronotuls • 
1.2,1• Borderline cases, bordering on 1olo. lexemes and syn- 
tagmas which cannot be identified morphologically/syn- 
tactically as proforms, but of which we know from 
experience that they do ocouz as proforms: Din~, tun. 
~eschehen, aus diesem Grund, etc. 
1.2o2o Semantic proforms per se: this group consists (mainly) 
of "synonyms', and, perhaps, all those lexemes which 
stand in a certain subconcept/superconcept relation to 
one another, ee g., like Engl, oar and vehicle. 
2. Semantic coherence without proforms. These are de- 
monstrated best b I an example: In 
(I) Das Haus gehoert mlr. Leider ist das Dach 
(This is my house• Unfortunately the roof is 
sohon baufaelllg, 
in bad shape°) 
we cannot apply any of the abovementioned olassiflcat- 
ion criteria, although it is obvious that semantically 
Haus and Dac_._~h belong together• 
After having established this classification a matrix 
was developed which codifies all the available morphological 
~nd syntactical information, plus information delivered by the 
concept-role structure of the net. In the matrix we find Inform- 
ation such as whether a proform substitutes for a noun phrase, 
a dependent clause, an independent clause, a text, an adverb 
etc°, or a combination of them, which preposition to expect in 
an answer to a question using a specific interrogative pronoun, 
etc• 
It has already been established (Hobbs 1978) that the 
information contained in the morphological-syntactic part of 
the matrix is not sufficient to identify semantic coherence in 
- !73 - 
a clause or text. Even for the proforms of 1.1. it is not 
sufficient to rely on the gender of the preform to find the 
correct antecedent. In the exe~nple- 
(2) Der Fisch wird auf den Teller gelegt. Dann esse ich lhn. 
(The fish is put on the plate. Afterwards I eat it.) 
Identification via gender provides us with the information 
that either Fisc h or Teller may be the antecedent of lhn. 
Our concept-role structure, on the other hand, informs us 
that - given ordinary circumstances - only fish are to be 
eaten, whereas plates are not. This is to say that morpholog- 
ical and syntactical knowledge give us an indication where to 
look for semantic coherence in a clause or text (syntax as 
"coarse* semantic (pre-) orderingl), but that we need criter- 
ia from the semantic net in order to be reasonably certain 
of what the proform stands for, or where semantic coherence 
is to be found in case there are no proforms. 
A good example of how the system VIE-LANG works is 
delivered by those syntagmatic adverbial proforms which stand 
for sentences or texts and which belong to 1.2.1.: 
(3) Das Restaurant ist heute geschlossen. Aus diesem Grund 
essen wir zu Hause. (The restauraunt is closed. For this 
reason we have dinner at home.) 
On the semantic "surface" there is nothing which relates 
Grund (reason) to anything else in either one of the two 
clauses. However, from our matrix we do know that aus diesem 
Grund may belong to 1.2.1 and, moreover, that its antecedent 
can only be a main clause or a text which immediately preced- 
es aus diesem G~do The system also recognizes that Restau- 
rant and essen belong to the same semantic neighborhood. Thus 
it follows that Das Restaurant ist lheute geschlossen assume 
the role +CAUSE with respect to essen wir zu Hause. 
From these examples it should be obvious that semantic 
coherence according to 2, has to be based exclusively on the 
- 174 - 
concept-role structure of the net. F.i., in example (1) se- 
mantic coherence can be made clear only when we know that 
Hau_~s and Dac...~h belong together semantically. How can we achieve 
that? Some scholars assume that a kind of inferencing guides 
us from Hau_..~s to Dach. In our system the lexeme "Haus" would 
be connected with the concept BUILDING, in which there exists 
a role ROOF, which is the net-structure the lexeme *Dech ° 
would lead to. Instantiation of a role always comprises the 
instantiation of the concept it belongs to and the value- 
-restriction respectively. The relationship between "Haus" 
and "Dach" becomes clear by merely accessing the net, without 
any inferences at allw 

References

Braohman R.J.: A Structural Paradigm for Representing Know- 
ledge, Bolt, Beranek and Newman; Cambridge Ma, 1978. 

Hirst G. : Anaphora in Natural Language Understanding: A Sur- 
vey, Univ. of British Columbia TR-79-2! 1979. 

Hobbs, J.R.: Pronoun Resolution, City Univ. of New York, 
Research Rep. No. 76-I; 1976. 

Hobbs, J.R.: Coherence and Coreference, SRI International; 
1978. 
