Dealing with Cross-Sentential Anaphora Resolution in ALEP 
Thierry Declerck 
IMS, University of Stuttgart 
Azenbergstr. 12 
D-70174 Stuttgart 
thierry@ims.uni-stuttgart.de 
Abstract 
The experiments described here have been 
done in connection with the LS-GRAM 
project, which is concerned with the develop- 
ment of large scale grammars and thus fore- 
seen the coverage of "real life texts". But in 
order to deal with such texts, it is also nec- 
essary to process linguistic units which are 
larger than sentences. The resolution of cross- 
sentential anaphora is one of the problems we 
have to deal with, when we switch towards 
the analysis of such larger linguistic units. 
In order to propose an analysis of the cross- 
sentential anaphora, one has to be able to re- 
fer back to an antecedent, which is to be found 
in a preceding sentence. This will be done 
on the basis of an information-passing frame- 
work. Using also the simple unification tech- 
nique a resolution of the pronoun can then 
be tried out: parts of the content informa- 
tion of the pronoun are going to be compared 
(unified) with specific parts of the content in- 
formation of the (possible) antecedent. 
1 Introduction 
The experiments described below have been done 
in connection with the LS-GRAM project 1, which 
is concerned with the development of large scale 
grammars. The specifications of the project fore- 
seen the coverage of "real life texts", which have 
also been processed by a corpus analysis. The re- 
suits of the corpus analysis allowed us to determine 
a priority list of the linguistic phenonema to be de- 
scribed. And in order to deal with "real fife texts", 
it is also necessary to consider the processing of lin- 
guistic units, which are larger than sentences. And 
as it is well known, the interpretation of sentences 
embedded in larger units is often distinct from the 
one of sentences, which are standing on their own. 
1The LS-GRAM (Large-Scale GRAMmars for EC 
languages) project is funded by the CEC under the 
number LRE 61029. The examples and the grammar 
descriptions I am using are taken from the German 
grammar, see (Rieder & al. IAI). 
The resolution of cross-sentential anaphora is one of 
the problems we have to deal with, when we switch 
towards the analysis (or synthesis) of such larger 
linguistic units. In order to give a correct interpre- 
tation of the cross-sentential anaphora, one has to 
be able to refer back to an antecedent, which is to 
be found in a preceding sentence, and I am using 
the term information-passing exactly in this sense: 
some information about a possible antecedent must 
be stored in order to be passed on to following sen- 
tences and to allow the anaphoric link, if some of the 
subsequent sentences are containing an anaphoric 
pronoun. Using the simple unification technique, 
as for the processing of other linguistic phenom- 
ena within ALEP, a resolution of the pronoun can 
then be tried out: parts of the content information 
of the pronoun are going to be compared (unified) 
with specific parts of the content information of the 
(possible) antecedent. 
In the next section I will first show how larger lin- 
guistic units can be processed within the ALEP sys- 
tem. In a second section I will very briefly present 
a semantic framework, which introduces the idea of 
"information-passing" in order to cope with cross- 
sentential anaphora: the Dynamic Predicate Logic 
(DPL). In the last section I will show how a very 
preliminary and tentative implementation of this 
framework can be modelled within the ALEP for- 
realism. Even if this first implementation is some- 
how primitive, this will permit us to formulate some 
remarks about the allowed degree of modularity of 
grammar descriptions within ALEP and also about 
the way in which such descriptions can be extended. 
These are two important aspects if one considers the 
task of developing large scale grammars. 
2 The Text Handling System and 
the ~Paragraph unit ~ 
Everyone who writes a grammar within the ALEP 
platform has some 'contact' with its Text Handhng 
(TIt) system, which converts each input into a 
SGML tagged expression. The TH component is 
the first processing step provided for by the ALEP 
system. In this tool, the sentence is defined as 
280 
tile default linguistic unit. If larger units are to 
be processed, this has to be explicitely defined by 
the user. lit our case, the linguistic unit is defined 
to be 'P' (for 'Paragraph') 2. The output of the TH 
being for example: 
<p> 
<s> 
<W>aohn</W> 
<W>sleops</W> 
<PT>• </PT> 
</s> 
</P> 
one can refer to the tag 'P' in order to define this 
structure being the linguistic unit to be processed. 
As usual (and also obligatory) for tile development 
of grammars within ALEP, a so-called ts_ls_rule 
(a mapping between text structures and linguistic 
structures) has to be defined: 
ts_is rule( 
id: { 
syn => syn:{ 
constype => phrasal:{ 
max => yes, 
constr => paragraph} } } }, ,p,, 
\[\] ). 
where a linguistic description ('ld') defining the 
'constr(uction)' type of a 'paragraph' is associ~ 
ated with the tag 'P', symbolizing the text type 
'paragraph'. The distinguishing value here is 'para- 
graph', which has been added to the type .system as 
a possible value for the feature 'constr': the ALEP 
formalism being type based, every feature, with its 
range of possible values, has to be declared in the 
declaration component. 
The next step involves in the description of gran> 
mar rules whict, parse tile structure of a para- 
graph. The phrase-structure rule responsible for 
the building of the paragraph-structure is simple. 
The mother node simply allows a binary branching 
of two sentential daughters. A recnrsion is defined 
on the right daughter, the value of 'constr' being 
a disjunction of 'punct_att' (describing a sentence 
terminated by a fall stop) and 'paragraph' (describ- 
ing thus the recursion). 'l'he left daughter is con- 
sidered to be the head (structure-sharing of 'head' 
features), as one can see in the following (simplified) 
presentation of the rule: 
id:{ 
syn => syn:{ 
constype => phrasal:{ 
constr => paragraph}, 
~I would like to thank Gordon Cruickshank ((\]ray 
Systems, Laxembourg) who gave me the initial idea to 
use this strategy in order to describe the interdepen- 
dency of information between sentences. 
category => head_cat:{ 
head => HEAD => v head:{} } } } } 
<\[ 
id:{ 
syn => syn:{ 
constype => phrasal:{ 
constr => punct_att}, 
category => head_cat:{ 
head => HEAD => v_head:{} } } } } , 
Id:{ 
syn => syn:{ 
constype => phrasal:{ 
constr => (punct_att ; paragraph) }, 
category => head cat:{ 
head => v_head} } } } \]. 
where '<' symbolizes the immediate dominance re- 
lation between the mother and the list of daughters. 
In principle , these are the steps which are neces- 
sary in order to extend the coverage of the gram- 
mar to larger linguistic units• There is naturally 
some more technical work to be done, but this will 
be described in the third chapter, where i will go 
into more details of the architecture of the gram- 
mar development. At this stage, we are able to 
parse a paragraph and to get a syntactical analyse 
of this structure. Sonic aspects which are specific to 
text linguistic should be considered. The one I am 
concentrating on is the cross-sententiM anaphoric 
relation. This has been postponed to the seman- 
tic which is treated within the re.finemenl compo- 
nent of the grammar. But before explaining the 
motiwLtion of the grammar design on this point 
and the reasons for postponing the semantic untill 
tile process of refinement, the semantic framework 
which has been choosen for the modelling of the 
cross-seutential anaphora should be presented• 
3 DPL as Representation Language 
for Information-Passing 
The Dynamic Predicate Logic (DPL) results from 
an investigation of a dynamic semantic interpreta- 
tion of the language of first order predicate logic 
and is "intended as a first step toward a compo- 
sitional, non-representational theory of discourse 
semantics ''3. This approach is concerned among 
other things with the cross-sentential anaphora. 
The dynamic aspect resides in the fact that, for 
this approach, the meaning of a sentence doesn't 
lie in its truth conditions, but "rather in the way 
it changes the ... information of the interpreter ''a. 
DPL considers only the information change which 
concerns %heir potential to 'pass-on' possible an- 
tecedents for subsequent anaphors"5. The Dynamic 
aI ant refering here to (Groenendijk91). 
'J Ibid. p. 43 
"~Ibid. p. 44 
281 
Predicate Logic is based on the syntax of the stan- 
dard predicate logic, but proposes a new (dynamic) 
interpretation of the quantifiers and connectives 
which allows the binding of variables within and 
outside their scope, depending on the interpreta- 
tion of the corresponding expressions of the natural 
language. 
Two (strong) assumptions, which are controversial 
in the discussion on this topic, are underlying the 
DPL approach: Indefinite NPs are considered to 
be quantifiea~ional expressions and pronouns to act 
like variables. Not everyone agrees on those as- 
sumptions, as this can be seen in the Discourse 
Representation Theory or in the work by Irene 
Helm 6. But those assumptions are here important 
if one wants to provide an uniform translation of in- 
definite NPs into existential quantifier (see below). 
And the desired compositional treatments requires 
that the information concerning the pronouns is to 
be found in the sentences uttered so far, i.e. as 
included within the scope of a logical quantifier or 
connective. 
The particular expressions of the natural language 
DPL is dealing with are the following: 
(1) A man walks in the park. He whistles. - cross- 
sentential anaphora 
(2) If a farmer owns a donkey, he beats it. - donkey 
sentence 
(3) Every farmer who owns a donkey, beats it. - 
donkey sentence 
And the problem consists in providing an ad- 
equate semantic representation of the anaphoric 
links. There are several ways of representing the 
semantic interpretation of each of the utterances 
and three of them (1 - 3) are discussed by Groe- 
nendijk & Stokhof: 
(A) In classical predicate logic: 
- ~x\[man(x)Awalk_in_the_park(x)Awhistle(x)\] (1) 
VxVy\[\[farmer(x) A donkey(y) A own(x, y)\] --~ 
beat(x, y)\] (2) & (3) 
(B) In a compositional way: 
- ~x\[man(x)Awalk_in_the_pavk(x)\]Awhistle(x)(1) 
- -~x\[farmer(x) A 3y\[donkey(y) A own(z,y)\]\] -~ be t( , y) (2) 
A  y\[donk y(y) A y)\]\] 
 e t(x, y)\] (3) 
(C) In the Discourse Representation Theory: 
- Ix\]\[man(x), walk~n_the_park(x), whistle(x)\] (1) 
\[_\] \[Ix, y\] \[farmer(x), donkey(y), own(x, y)\] --~ y)\]\] (2) (3) 
What is missing in (A) is the compositional repre- 
sentation of the subparts of the utterances. Another 
disturbing point is the distinct translation of the 
6See (Kamp 1981) and (Heim 1982, p. 122) 
indefinite NPs into the representational language, 
once as an existential quantifier (A,1) and once as 
an universal quantifier (A,2 & A,3). The fact that 
(2) and (3) translate into the same semantic repre- 
sentation is also reflecting the non-compositionality 
of the classical predicate logic in this case. 
The problems with the compositional representa- 
tion (B) are concerning the binding of the variables 
(the pronouns in the natural language). In (B,1) 
the third occurence of the variable x is free and thus 
doesn't allow the anaphoric reading. The same re- 
marks are valid for x and y in (B,2) and for y in 
(B,3). But the way (S) is representating the utter- 
ances allows the uniform translation of indefinite 
NPs into an existential quantifier. 
The problems with the DRT representation are 
more of methodological nature, since on the treat- 
ment of those cases, DPL and DRT are empirically 
equivalent. In short: Groenendijk and Stokhof are 
missing the compositional building of the seman- 
tic representation and also would prefer to use a 
more classical representational language, like the 
one of first order logic. For this, they are 'merg- 
ing' together the representation (A) and (B), and 
considering now only the first case (1), the dynamic 
semantic interpretation is going to be like (B,1): 
~z\[man(x) A walk_in_the_park(x)\] A whistle(x), 
but with the existential quantifier having scope over 
the conjunction of the two sentences, this represen- 
tation is going to be equivalent to: 
-~x\[man(x) A walk_in_the_park(x) A whistle(x)\]. 
This is possible because the interpretation of a sen- 
tence doesn't lie in a set of assignments, but rather 
in a set of ordered pairs of assignments, where those 
pairs represent the input-output states of a sentence. 
In our example, the first sentence has an output 
which is as the same time the input of the second 
one. Since the existential quantifier is interpreted 
as being able to quantify outside its scope (also in 
combination with the conjunction and the sequenc- 
ing of sentences), the information concerning the 
(possible) antecedent is going to be passed-on to 
following sentences, which could be subsequently 
uttered. The fact that the existential quantifier 
in DPL is interpreted as a quantifier which can 
bind outside of its syntactic scope allows to say 
that we provide a compositional treatment of the 
utterance, the second sentence being interpreted as 
it comes, without referring to some metalinguistical 
representation or process. The existential quantifier 
is qualified as an externally dynamic quantifier. 
Not every quantifier (or connective) has the dy- 
namic property of binding outside of its scope; the 
universal quantifier, for example, can bind within 
its scope, but not outside of it: 
(4) *Every man walks in the park. He whistles 
282 
is ruled out. The dynamic semantic interpretation 
of this quantifier blocks the passing of the informa- 
tion: the ouptut of the first sentence is empty (with 
respect to the information concerning anaphoric 
binding). The input of the following sentence will 
therefore contain no information allowing a resolu- 
tion of the pronoun. 
The way DPL is interpretating the distinct quanti- 
tiers and connectives is the following one: 
• Existential quantification and conjunction are 
e;ctcrnally dynanfic. 
They can bind variables within and outside 
their scope: 
\[A man\],/ walks in the park and hei whistles. 
IIe,i is happy 
• Universal quantification and implication are 
internally dynamic. 
They can bind variables only inside their scope: 
Every farmer who owns \[a donkey\]i, beats it,i 
• \[Every man\],i walks in the park. He/ whistles 
If \[a farnrerJ~ owns \[a donkey\]j, he~ hates itj 
• If \[a farmer\]~ owns \[a donkey\]j, he beats it. 
He,: hates it i 
• Negation and disjunction are static. 
They cannot bind .variables (at least, they 
don't allow a anaphoric reading): 
• \[No man\]~ walks in the park. tte~ whistles 
• \[A man\],i walks in the park or her whistles 
This is too simple and for sonre English examples 
it seems to be wrong. The authors arc consider- 
ing and discussing the cases which contradict the 
assumptions and give some hints in order to inte- 
grate those cases. I will not discuss this point here, 
but just mention, that for the German grammar 
we should have a look at a detailled analysis of the 
meaning of such expressions 7. Once this has been 
done, we can encode this information in the lexi- 
con (as will be seen in the next section). But here 
we can say that the DPL approach allows us, to 
a certain degree, to account for the resolution of 
anaphora without having to leave the field of lin- 
guistic descriptions. With the only means of the 
grammar and the formalism we have, we are able 
to provide a tirst and simple description of those 
phenomena. It is still to be investigated how so- 
phisticated such a treatment can be. 
4 A first Implementation of the 
Dynamic Interpretation 
As we have seen, the 'paragraph' has been defined 
as the linguistic unit to be processed by the sys- 
teni. To provide a (simple) syntax was so far not 
a problem. But, as stated in the second section, if 
~As for example in (Bethkc 1990) or in (Vater 1979). 
a free pronoun occurs in a sentence, it is possible 
that this pronoun requires an anaphoric interpre- 
tation. To achieve this interpretation, some infor- 
mation about the antecedent is necessary and this 
information is to be found in a precedent sentence. 
DPL theory provides us with an elegant framework, 
describing the semantic of utterances as the way in 
which information is passed-on between sentences 
and so controlling the possible binding of pronouns. 
4.1 The Organisation of the Lingware 
I have tried to model the DPL framework within the 
ALEP platform. This experiment is documented in 
the following section. In doing this, I followed the 
overall strategy of the grammar development within 
ALEP. The syntactic 'paragraph' rule has been de- 
scribed within the analysis component of the gram- 
mar. The process of analysis is a process concerned 
with the building of structure trees induced by the 
ps-rules. In our grammars, this process is associ- 
ated with a subpart of the lexicon, which is con- 
taining only the information relevant for that kind 
of process, i.e. the building of a parse tree, which is 
only dealing with morpho-syntactical information. 
One of the motivation of this organisation of the lex- 
icon (and also of the rules associated with it) lies 
in tile consideration of efficiency. At least for one 
of the parser of the system (the bottom-up head- 
out parser), the presence of nmltiple entries for one 
item and the description of more than one rule for a 
phenomenon has very negative consequences on the 
run-time behaviour of the system, backtracking be- 
ing very expensive. But even if another parser (the 
record parser, which is not so sensitive to this kind 
of problems) is used, the use of shallow linguistic 
descriptions in the analysis component allows one 
to formulate some generalizations. Lexical and se- 
mantic ambiguities are than resolved or introduced 
at the following level, the refinement component, 
which corresponds to a process of decorating al- 
ready existing trees: feature structures are just 
added to the trees. The refinement process is a spe- 
cial feature of the ALEP formalism. This process 
consumes very little cpu-time. For this process, spe- 
cific linguistic subcomponents are described. 
There is a relation between the sublexicons, this 
relation being one of subsumption. Linguisitic de- 
scriptions contained in the distinct sublexicons have 
all the same structure but the degree of specification 
is different from one lexicon to another. Thus we 
don't have a stratificational model and the descrip- 
tions stay declarative and monotonic. We are just 
describing what subparts of information of an entry 
is going to be accessed by a process. Since this or- 
ganisation of the lexicon (and the associated rules) 
is done along the line of the processes provided by 
ALEP, we call it the vertical organisation. 
The description of this vertical organisation is done 
with the help of the 'specifier features' provided 
283 
by the ALEP system, wich are configurable by the 
user. It is in such a way possible to determine what 
subparts of the grammar are going to be accessed 
by a particular process. 
4.2 The Lexicon Entries and the Rules for 
the Resolution of Anaphora 
The rule building the paragraph structure intro- 
duced before produces just one tree. The possi- 
ble distinct readings of it are described within the 
refinement conlponent. Thus the cross-sentential 
anaphora relation (being essentially a semantic 
process) is fully described on this level (the morpho- 
syntactical aspects being described in the analysis 
coniponent of the grannnar). 
We will no see how the relevant items are described 
in the ~'e\]inc lexicon. I just consider here the entries 
of substantives, pronouns and quantificational ex- 
pressions (determiners). My goal is to provide the 
informations that are necessary for the modelling 
of the dynamic treatment of the cross-sentential 
anaphora. Here the way they arc coded in the Ger- 
lnan gralnlnar: 
• Referential expressions 
ld:{ 
content => ... 
restr => 
\[inst_zero_psoa: { 
indx 
tel => tel:{ 
rel_name => flugzeug }, 
inst => A48 } \], 
=> A48 => ind_indx:{ 
pets => p3, 
numb => sing, 
gend => neut } 
}}}. 
• Pronominal expressions 
id:{ 
content => ... 
restr => 
\[inst_zero psoa:{ 
rel => tel:{ 
rel name => PRO }, 
inst => A48 } \], 
indx => A48 => ind indx:{ 
pets 
numb 
gend 
}}}}. 
• Quantificational expressions 
dot_d_das " 
mLDref_core \[ 
det d das, 
das, 
mLA sem_funct_sem_det\[_\] \] . 
=> p3, 
=> sing, 
=> neut } 
det_ein" 
mLDref_core \[ 
det_ein, 
ein, 
mLA_sem_funct_sem_det \[extern_dynamic\] \] . 
det_kein ~ 
mLDref core\[ 
det_kein, 
kein, 
mLA_sem_funct sem_det \[static\] \] . 
det_d_jedes ~ 
mLDref_core \[ 
det_j edes, 
jedes, 
mLA_sem_funct_sem det \[intern_dynamic\] \] . 
I encoded the infornlation about semantic gender 
and number of the referential expressions and the 
pronouns (contained in the 'restr(iction)' attribute. 
The 'rehname' attribute represents the referential 
property of the item: this a variable ('PRO') in 
the case of the pronouns, nlodelting in this way the 
DPL assumption that pronouns are acting as vari- 
ables. The entries of quantifiers and determiners 
are presented here as macros. The relevant in- 
formation for us is the one concerning the quan- 
tificational force of this entries (I don't consider 
here binary conjunctives) which is lexically deter- 
mined. In case of 'das' (the), the relevant informa- 
tion has been left unspecified, since where are not 
considering definite descriptions for the time being. 
The quantificational force of 'ein' (a, an) has been 
specified as 'extern_dynamic', the one of 'kein' (no, 
no one) as 'static' and the one of 'jedes' (every) as 
'intern_dynamic', nlodelling the classificiation pro- 
posed in the DPL framework. 
During the processing of the paragraph, if a free 
pronoun occurs in a sentence and if it can refer to an 
antecedent, the value of the 'quantificationahforce' 
of the antecedent should be 'extern_dynamic' and 
the values of the attribute 'restr' of the antecedent 
and the pronoun must be unifiable. In this case, 
the value of the 'rehname' of the pronoun is uni- 
fied with the value of the one of the antecedent, 
as one can see below, where in this case the val- 
ues of the 'restr' features are variable-shared. Here 
a (simplified) rule accounting for resolution in the 
context of one-argument predicates (the other cases 
are described by rules disjunction): 
ld:{ 
sem => sem:{ 
content => lq_cont:{ 
argl => lq_cont :{ 
quants => \[quantifier :{ 
q_force => QI } \], 
rd_cont => r indx:{ 
restr => RESTR1 } } }, 
284 
argl => lq_cont:.\[ 
quants => \[quantifier:{ 
q force => ~2 }\], 
rd cent => r_npro:{ 
restr => RESTR2 } } } \] 
7}}} <\[ 
id:{ .... 
sere => sore:{ 
content => lq cent:{ 
argl => lq_cont:{ 
quants => \[quantifier:-\[ 
q force => ~i } \], 
rd_con% => r npro:-\[ 
rostr => RESTRI } } } } } \] }, 
id:{ .... 
som => som:{ 
content => lq cent:{ 
argl => ( lq cent:{ 
quants => \[quantifier:{ 
q_force => ~1 => extern dynamic} \], 
rd_cont => r ppro:{ 
restr => RESTRI } } 
/ quants => \[quantifier:{ 
q_force => Q2} \], 
rd_cont => r_npro:{ 
restr => RESTR2 } } } } } } } \]. 
With this simple technique, we are able to ac- 
cept "Ein Mann kommt. Er singt" and to reject 
"Jeder Mann kommt. Er singt .... the value of 
'quantifieational._force' of the entry 'jeder' (every) 
is 'intern_dynamic' and so disallows the unification 
of the 'restr' wdnes. The same with "Kein man 
kommt. Er singt". The negation is a static se- 
mantic phenomenon and the value of its attribute 
'quantifieational_forcc' is 'static'. a If there is no 
pronoun in the second sentence, no unitication is 
tried out: the values of the corresponding q'cstr' 
attributes arc not structure-shared. The building 
of the paragraph structure just goes on. The values 
of the. attributes 'quantificational_foree' and 'restr' 
are then put together in a list. 
4.3 Modularity and Exteudability 
The short experinrent described in this paper is re- 
ally too primitive and doesn't allow any statement 
about the possibility of providing a conrplete treate- 
ment of the cross-sentential anaphora on the basis 
of the DPL framework. But one goal of the exper- 
iment was also to gain some knowledge about the 
possible extension of the coverage of the grammar. 
And heresome conclusions can be drawn. 
8Linguisitically speaking, this is too simple. In Ger- 
m~m at least, negated NPs can often bind pronouns. 
And we should also allow generic readings. Stone ex- 
periments have been done with respect to this. But ac- 
tuMly, I just would like to show how information-passing 
can be modelled in ALEP. 
First of all, the modular organisation of the gram- 
mar development within ALEP proved itself to be 
very practical. The 'tools' provided by ALEP (the 
'specifier features'), if they are reasonably config- 
urated during grammar development, allow a high 
degree of modularity and pernlit, without difficul- 
ties, to define new grammar components. 
With the help of the text handling component, it 
was also no problenr to extend the coverage of the 
grammar to larger linguisitc units. The descrip- 
tion of grammar colnponents for such units is well 
supported. 
It still remains the task of providing some prefer- 
ence descriptions for anaphora resolution. This will 
1)e done Mong a more detailled linguistic analysis 
and considering corpus analysis. ALEP provides 
also for constraint solvers which allow to define such 
preferences in an elegant (propositional) way. 
RcfercIlce8 
BIM/SEMA, ALEP System Docume'atatiou, CEC 
1993, l,uxembourg. 
Simpkins, N.K.; Groenendijk, M. and Cruickshank 
G. (P-E International), ALEP 1 UseT, Guide, CEC 
1993, Luxembourg. 
Badia, '\['.; Bredenkamp, A.; Declerck, T.; Itentze, 
R.; Marimon, M.; Schmidt, P.; Theofilidis, A. 
(1995), LS-GRAM Rule Coding Manual, Deliver- 
able D - WP7 Version 1, CEC, 1995, Luxcmbourg. 
Bethke, I. dc'r die das als Pronomen, 1990. 
J. Groenendijk & M. Stokhof, Dynamic Predicate 
Logic, in Linguistics and Philo.sophy 14, p. 39-100, 
1991 
Ileim, Irene, The. Sema.'l~th:s of Definite. and \]'a- 
definite Nou'a Ph.r'ases, dissertation, Uifiversity of 
Massachussetts, Anrherst, 1982. 
Kamp, itans. A Theory of Truth and Semantic Rep- 
resentation. In 3. Groenendijk, T. Janssen, and M. 
Stokhof (eds.), Formal Methods in the Study of Lan- 
guage,, Mathematical Center Amsterdam: 277-322, 
1981 \[reprinted in .1. Groenendijk, T. 3anssen, and 
M. Stokhof (eds.), Truth, I',,ter'prctation and furor 
ranlion, Foris, Dordrecht, 1-41, 1984. 
Sibylle Rieder, Paul Schnfidt & Axcl Theofil- 
idis- IAI, & Thien'y Declerck - IMS. LS-GRAM 
Ling'war('., Docmnentation, Deliverable C-D-WI?6e, 
1996, Luxcmbourg. 
Vater, Hein~. Das System de'r Artikelforme'o. im 
g('.gcnwae.r~igen De'~Ltsc.h.. 2. edition, ~lhlebingen, 
1979. 
285 
