'HOW "/0 RESTRICT AMBIGUITY 0F DIBCOURSE 
Barbara Dunin-K@plic z 
Institute of Informatics 
University Of Warsaw 
P.O. Box 1210 
00-901 Warszawa, POLAND 
ABBT~CT 
We single out a class of prototypes i.e., 
a class of constructions forcing the obligatory 
coreference or obligatory noncoreference. An 
essential feature of prototypes is their 
undistinctiveness. In this sense they are the 
most natural and efficient mearis of 
communication in discourse. 
The non-application of prototype should be 
well motivated. This leads to the rule of 
restricted choice stating that whenever it is 
possible the application of a prototype should 
be preferred. 
The rule of the restricted choice suggests 
the general outline of interpreting ambiguous 
sentences, strictly speaking, the method of 
ordering admissible interpretations: those which 
can be equivalently expressed by means of 
a prototype are less probable. In other words, 
the rule of the restricted choice can be 
regarded as some kind of mechanism ordering 
the hypotheses for computer/on. 
INTRODUCTION 
The crucial problem in discourse analysis 
is the appropriate transposition of all 
expressions occurring in it, into reality (see, 
for instance, the framework provided by Kemp 
in (Kamp, 1981)). Even preliminary analysis 
shows that one real object can be identified 
by various surface constructions. This forces 
the necessity of dividing surface expressions 
into classes denoting identical individuals. 
The above problem can formally be stated 
as follows. To each discourse D we assign 
some reality which can be understood as 
a set ID of individuals (semantic discourse 
domain) together with a set of relations 
defined on ID. The semantic discourse domain 
can be interpreted twofold: 
1.o as a set of real objects i.e., objects 
existing in actual world; 
2 ° as a set of mental objects i.e., objects 
existing in language user's mind. 
Although the first interpretation is more 
natural, it leads to some ontological problems, 
concerning the distinction of fictitious and 
non-fictitious entities. Since there is no such 
distinction from linguistic perspective the 
second interpretation seems better motivated. 
"/he next step is to define syntactic 
discourse domain, denoted by S_, i.e., a set L) 
of discourse expressions referring to 
individuals (set of individuals). The mapping 
assigning individuals to syntactic expressions 
will be ca/led the reference function and 
denoted by R. F'orma/ly, R : S D 2 ID. 
Example 
(DI) John and Peter admire winter. They 
are often skiing together. 
SDI" - {"John", "Peter", "winter", "they '''~ 
ID 1 - ~John, Peter, winter~ 
("John") - {John} 
R ("Peter") - {Peterl 
R ("they") ,, {John, Peter\]" 
R ("winter") - {winter} 
We say that discourse expressions x and 
y a.re coreferencial, what we denote by xCy, 
if and only if they refer to the same set of 
individuals. 
Formally, 
for each x,y ~ S u xCy iff R(x)= R(y) 
It is readily verified that C is an 
equivalence relation. Obviously each 
equivalence class of C contains coreferentia/ 
expressions. The set of equivalence classes 
of C will be ca/led the reference sta~te of 
discourse and denoted by RSD. 
Example 
(D2) John took a knife. 
( RSD 2 ) " "~ knife '~-. John 5~ ~"a 
(D3) John took a knife. He hurt himself. 
(RSD3) ~"JOhn:', "he", "himself'.~ knife"}. 
We can observe here that adding new 
utterance to the discour.~;e changes its 
93 
reference state. In this sense RSD is a 
dynamic notion Let us note also that the 
problem of anaphora solution can be regarded 
as defining the relation C for the whole 
discourse. 
Both the speaker, while constructin~ 
a discourse, and the hearer, while eunalysing 
it, try to achieve the identity of RSD at each 
step of the discourse. We argue in this 
paper that to accomplish this effect, the 
speaker has at his disposal (at each moment) 
a more restricted set of linguistic 
constructions than it seems intuitively. Let 
us notice that expressions belonging to one 
equivalence class have various syntactic 
shapes at different steps of discourse. It ca/'* 
be shown that the syntactic form of 
expressions at particular moments is not 
accidential, i.e., elements of indicated 
equivalence class are not interchangeable. 
P R O'I'O'I'YP E S 
Recent discourse theories provide several 
levels of larlguage analysis: morphological, 
syntactic, semantic and sometimes pragmatic. 
Each of these levels determines a 
characteristic set of notions and mechanisms. 
It is assumed here that the analysis of 82", 
utterance on each levels of language should 
yield complete information obtainable by tools 
available on this level 
Classical anaphor resolvers act on 
semantic level on discourse analysis. ~Are take 
the position that for inflexion al languages the 
coreference relation can be partially described 
on the syntactic level An essential feature of 
this partial characteristic is defining the 
coreference relation quite apart from Peal 
individuals, i.e. without specyfing the reference 
function. 
"Po fix some ideas let us consider an 
utterance containing the noun phrases NPl, 
..., NP . If there is no information regarding n 
coreference, all we can do is to assert that 
the coreference relation is included between 
the "minimal" relation, i.e., relation identified 
by the unit equivalence classes NP~ , ..., 
NP and the mcLximal one, i.e. adm,Ring in 
one nclass all number-gender agreeable 
phrases. 
We very seldom deal with such a situation 
in practice. Almost a/ways we can assign to 
an utterance a syntactic level information 
stating obligatory- coreference or obliqatory 
noncoreference of some expressions. 
The surface constructions carrying this 
kind of information with respect to pronouns 
and zero pronouns (in the case of elided 
subject) will be called prototypes. Ln other 
words prototypes can be regarded as 
syntactic means forcing obligatory coreference 
or obligatory noncoreference between pronouns 
or zero pronouns and other surface 
expressions. 
Let us consider few instances of 
prototypes. Because the ideas presented here 
are implemented for the Polish language, the 
notion of prototype will be illustrated with 
a number of Polish sentences. An elided 
subject specific for inflexional languages can 
be observed here. It is clenoted by (~ 
Because elided subject expresses some 
aspects of thematic continuity, its 
interpretation seems to be an important step 
during discourse analysis. English 
translations of presented examples preserve 
their syntactic shape. Unfortunately they are 
sometimes incorrect as English sentences. 
(1.) Piotr obudzit si~1 ~)3.. podszec£~ do okna, 
~2 ot~vorzy~ je i ~)3 wyskoczy{. 
Peter woke up, ~I came to the window, 
 2 opened it and ~)3 jumped out. 
Expressions: 
Peter, #I, ~2, ~)3 are coreferentiaL Another 
interpretation is unadm,ssible, in (I) we deal 
with obligatory coreference of expressions 
(denoted by a - -- b). 
(2) ~I~2 podszeci% do okna, 
4~4 ~skoczy~ 
~)1. Woke up, q)2 came to the window, 
q)3 opened it and #4 jumped out. 
in (2), similarly as in (a) (co-ordinate 
clauses) and in (3), (4) (subordinate 
clauses) the only acceptable interpretation 
is explicitely showed. 
(3) Z~im Ca,~s~ea~2 zgasi~ ~wiat~o. 
Before 911 left, 912 turned the light off. 
(4) ~\].~_Z_gasi~ ~wiat~o, zanim....w~2 wyszedL 
911 Turned the L~ght off, before ~2 left. 
The next examples concern the obligatory 
noncoreference of expressions (denoted by a+-b) 
(5) Ona lubi ja~ 
She likes her. 
(6) (~ zapyta~ Piotra,'~czy Jan p6jdzie 
do teatru. 
¢ Asked Peter, whether John would 
go to the theatre. 
94 
(7) ¢ Usiad~ do stc~u, a .Tan naleuI: mu wina. 
Sat at the table, und John poured 
him out some wine. 
The above examples pose the question of 
how the class of prototypes should be singled 
out. This problem can be solved by 
specifying a collection of rules concerning 
the obligatory coreference and obligatory 
noncoreference. The exact format of these 
rules is beyond the scope of this paper. 
For inflexional languages they depend on the 
type of considered sentence, the sentence- 
-level functions of considered phrases 
and their syntactic shape. As a simple 
example of such a rule let us consider the 
basic criterion of excluding coreference: 
If the object is expressed by means of 
a reflexive pronoum, then it is coreferential 
with the subject; in other cases the 
referential identily of the subject and object 
is excluded. 
This criterion can be applied both for 
deterrninig coreferents of objects - blocking 
the subject, and in tesf/n~ the possible 
antecedents of the subject - blocking the 
objects. This is exactly the case we have in 
(5). 
THE RULE OF' RESTRICTED CHOICE 
A conclusive criterion of being a prototype 
results from analysing a given sentence it% 
isolation. If it is possible to assert or to 
exclude the referential identity of some 
expressions of the sentence, indepedently of 
its context then the sentence can be regarded 
as an instance of prototype. An essential 
feature of prototypes is that they are 
completely indistinctive and in this sense they 
are the most proper tool for expressing 
a certain relationship in the utterance. This 
strong relationship makes it possible to 
eliminate some interpretations, which in other 
cases should be regarded as probable too. 
If we agree that achieving unambiguity of 
discourse is the major goal both for the 
speaker and the hearer, then the non- 
-application of prototype, as the most natural 
and efficient mea/qs of communication should 
be well motivated. When such a special reason 
is lacking, the speaker should apply a 
a prototype. Under this assumption the set of 
linguistic tools available to the speaker is 
restricted. 
The notion of prototype can be natural/y 
applied on the syntactic level of discourse 
anadysis to limit the number of hypotheses for 
further consideration. 13ut it can also be useful 
on the hi~her levels to interpret ambiguous 
discourses. Strictly speakin~ the properties 
of prototype suggest the general outline of 
interpreting ambiguous sentences, more 
precisely a method of orderin~ possible 
interpretations with respect to their plausibility. 
From the set of possible interpretations of 
a sentence, those that can be equivalently 
expressed by means of prototype, should be 
regarded as less plausible. The justification 
of this choice is clear: if the speaker wanted 
to point out such an interpretation, he would 
naturally achieve it by applying a prototype. 
In view of the obove we can formulate 
the rule of restricted choice. It states that 
whenever it is possible the application of 
a prototype should be preferred. 
It is irrportant to notice that the rule of 
restricted choice can be viewed from the 
perspective both of the speaker contructing 
the discourse and the hearer modelling it. 
The speaker should apply prototypes 
whenever it is possible. The hearer should 
take this fact into consideration. 
Let us try to interpret the concrete 
sentences with the help of the rule of 
restricted choice. 
(8) Zanim ~)I wyszed~, .Tan zgasi~ ~wiat{o. 
Before ~I leftmasc, .Tohn tumedmasc 
the light off. 
There are two interpretations here: 
(9) Zanim zgasi   wia o 
Before ~I left, John turned the light off. 
l 
(1.0) Zanim ~I w3zszed2, Jan zgasi{ ~wiat~o. 
Before ~ll left, John turned the light off. 
(~ denotes the reference to the context). 
But the first interpretation can be expressed 
by means of prototypes. 
(Pl) Zanim ~I ~Aryszed{, @2 zgasit ~wiat~o. 
Before ~1 left, ~2 turned the Light off. 
(P2) Jan zanim wyszedt, ~1 zgasit gwiatto. 
John before left, ~1 turned the light off. 
According to the rule of restricted choice 
the first interpretation should be regarded as 
less probable (note that it can be expressed 
by prototype). Hence the second interpretation 
should be chosen. 
9S 
Another example is more complicated. 
(11) Zanim o__.nn wyszed~, ~I zapy~ak ~eo, 
czy #2 p6jdzie do kina. 
Before he leftmasc , @I, had asked him, 
whether ~2 would have gone to the 
cinema. 
In the embedded clause 
03. zapy%a~ gO 
there acts the rule of obligatory noncoreference 
excluding the referentia/ identily of subject and 
object in this sentence: 
Zanim on wyszed~, @~zapytakj~o, 
czy ~2 p6jdzie do kina. 
According to our definition the above sentence 
is an instance of a prototype. 
Excluding the coreference of pointed 
expressions decreases the numer of possible 
interpretations, but does not clear up all 
referential relationships in this ambiguous 
sentence. Although there are no further 
syntactic premises to resolve this ambiguity 
we can specify the less probable interpretation 
by appying the rule of restricted choice. If the 
speaker wanted to express the following 
sense: 
(12) Zanim X wyszeck%, X zapy%a.% go, 
czy @1. p6jdzie do kineu . 
he should have used the following (structural) 
prototype: 
(13) Z~im ¢~1 ~apyt~ go, 
czy ~2 p6jdzie do kina. 
(In such a sequence of clauses in the 
sentence the rule of obligatory coreference 
demandes that ¢ and @1. shou/d be identified). 
It follows therefore that the interpretation: 
(3.4) Zanim on wyszedku.~l zap vta~ go, 
czy ~2 p6jdzie do kina. 
is the less probable and should be computed 
as the last one. 
NONMONOTONICITY OF THE RULE 
05' lqESTRICTED CHOICE 
Consider the following example: 
(15) I<iedy @I podszed~ do Piotra, 
by~ on zdenerwowany. 
When ¢I came near Peter, 
he was nervous. 
There are two possible interpretations 
(3.6) Kiedy ¢I podszed{ do Piofra, by~ on 
zden erwowany. 
( 1,7 ) Kiedy ~p.odszed~ do Piotra, by~A, on 
zden erwowany. 
2ecause the second interpretation can 
unambiguously be expressed by the prototypical 
c on struction: 
(3.8) Kiedy ~I podszed{ do Piotra, by~ ~)2 
zden er%vowalny. 
When ~)I came near Peter Q2 was 
nervous. 
according to the rule of restricted choice the 
first interpretaf/on should be preferred. 
The rule of resfx'icted choice is based on 
the assumpt/on that whenever it is possible 
people use unambiguous constructions. 
Although usually va/id this assumption cannot 
be regarded as genera@ truth. This meeuns that 
the rule of restricted choice enables one to 
jump to plausible but not ironclad conclusions. 
"l~pically, such conclusions can be invalidated 
when new information is available. In our 
example the preferred intezq0retation might he 
overturned when we extend our discourse as 
follows: 
(1.9) I<iedy ~I podszed~ do Piotra, 
by~ on zdener-vvowany. 
BM~ tc wynik wcze~niejszej k~6tni 
z Piotrem. 
¢I came near Peter he was When 
nervous. 
That was the result of an earlier 
quarrel with Peter. 
96 
The neoessity of changing the preferred 
interpretation follcws from the fact that new 
information is available. The property of 
drawing plausible but defeasible inferences 
characterizes non-monotonic reasoning. 
Various forms of this kind of reasoning are 
now being developed (see (~AI-84)). 
It is now widely recognized that discourse 
understanding regulres nonmonotonic 
mechanisms in many aspects. The rule of 
restricted choice is an example of such 
a nonmonotonic tool 
C ONCLUSIONS 
(1.) While constructing discourse the 
speaker wants the hearer to understand him 
correctly. Even if he uses ambiguous 
constructions he intends to cemrr, unicate the 
unique interpretations, and not to create in 
heater's mind a set of a/l possible hypotheses. 
It follows that constructhag NLU systems, which 
~enerate all admissible interpretations, 
contradicts common sense reasonln~. So the 
essential problem is to determine methods o/ 
choosing the most appropriate interpretation. 
If this plausible interpretation fails, it should 
be revised. 
(2) Employing the rule of restricted choice 
assumes the existence of some mechanism 
which determines whether a given construction 
can be regarded as a prototype. 'l~bis can be 
achieved by specifyins a set of rules quali~j'ing 
the obUgatory coreference and noncoreference 
of referrins expressions. A partied set of such 
rules for the l::ollsh language has been 
presented in (Dunin-K~plicz, 1983). 
ACKNOWLED GEMEN'P 
I would llke to thank Dr Witold Lukaszewicz 
for helping me to clarify ideas presented in 
this paper. 
REFERENCES 
Dunin-K@plicz B. (1983) Towards better 
understanding of anaphora, in: Prec. of 
the ist ACL Conference, Piss, 139-144. 
Kamp H. (1981) A theory of truth and 
semantic representation, in: 5"ormal 
methods in the study of language, 
IV\[athematisch Centrum, Amsterdam, 
277-322. 
AAAI (1.984) Non-monotonic reasoning 
workshop, New York 0 
97 
