ParseTalk about Sentence- and Text-Level Anaphora 
Michael Strube and Udo Hahn 
C~Z~- Computational Linguistics Research Group 
Freiburg University 
D-79085 Freiburg, Germany 
{strube, hahn}@coling.uni-freiburg.de 
Abstract 
We provide a unified account of sentence-level 
and text-level anaphora within the framework of 
a dependency-based grammar model. Criteria for 
anaphora resolution within sentence boundaries 
rephrase major concepts from GB's binding theory, 
while those for text-level anaphora incorporate an 
adapted version of a Grosz-Sidner-style focus model. 
1 Introduction 
This paper treats the resolution of anaphora with- 
in the framework of Parse Talk, a dependency- 
oriented grammar model that incorporates strict 
lexicalization, head-orientation (based on valency 
specifications), feature unification, and inheri- 
tance among lexicalized grammar specifications 
(Br6ker et al., 1994; Hahn et al., 1994). The re- 
sults we present rest upon two major assumptions: 
1. As many forms of anaphors (e.g., nomi- 
nal and pronominal anaphors) occur within 
sentence boundaries (so-called intra-sentential 
or sentence anaphora) and beyond (inter- 
sentential or text anaphora), adequate theo- 
ries of anaphora should allow the formulation 
of grammatical regularities for both types us- 
ing a common set of grammatical primitives. 
2. Anaphora are only one, yet very prominent 
phenomenon that yields textual cohesion in 
discourse. Adequate grammars should there- 
fore also be easily extensible to cover non- 
anaphoric text phenomena (e.g., coherence re- 
lations, rhetorical predicates), which provide 
for additional levels of text (macro) structure, 
with descriptions stated at the same level of 
theoretical sophistication as for anaphora. 
First, we will briefly compare our approach with 
work done in the context of government-binding 
(GB) grammar and discourse representation the: 
ory (DRT). As we conceive it, binding theory as 
developed within the GB framework (Chomsky, 
1981; Kuno, 1987) offers one of the most sophis- 
ticated approaches for treating anaphora at the 
sentence level of description. This has also been 
recognized by advocates of competing grammar 
formalisms, who have elaborated on GB's bind- 
ing principles (cf., e.g., Pollard and Sag (1992) 
within the context of HPSG, whose treatment is 
nevertheless restricted to reflexive pronouns). In- 
terestingly enough, when faced with some cru- 
cial linguistic phenomena, such as topicalization, 
GB must assume rather complex movement oper- 
ations in order to cope with the data in a satisfac- 
tory manner. Things get even more complicated 
when languages with relatively free word order, 
such as German, are taken into account. Finally, 
considering the case of text anaphora, binding the- 
ory has nothing to offer at all. 
Another strong alternative for considering ana- 
phora constitutes the framework of DRT (Kamp 
and Reyle, 1993). Its development can be con- 
sidered a landmark in the model-theoretic se- 
mantic analysis of various forms of quantified 
sentences, conditionals, and anaphorically linked 
multi-sentential discourse. At this level of descrip- 
tion, DRT is clearly superior to GB. On the other 
hand, its lack of an equally thorough treatment of 
complex syntactic constructions makes it inferior 
to GB. These deficits are no wonder, since DRT is 
not committed to any particular syntactic theory, 
and thus cannot place strict enough syntactic con- 
straints on the admissible constituent structures. 
Focusing on the text analysis potential of DRT, its 
complex machinery might work in a satisfactory 
way for several well-studied forms of anaphora, 
but it necessarily fails if various non-anaphoric 
text phenomena need to be interpreted. This is 
particularly true of conceptually-rooted and prag- 
matically driven inferences necessary to build up 
textual macro structures in terms of coherence 
relations (Hobbs, 1982) or rhetorical structures 
(Mann and Thompson, 1988). This shortcoming 
is simply due to the fact that DRT is basically a se- 
mantic theory, not a comprehensive model for text 
understanding; it lacks any systematic connec- 
237 
tion to comprehensive reasoning systems covering 
the conceptual knowledge and specific problem- 
solving models underlying the chosen domain. 
Summing up, DRT is fairly restricted both with 
respect to the incorporation of powerful syntactic 
constraints at the sentence level and its extension 
to the level of (non-anaphoric) text macro struc- 
tures. GB, on the other hand, is strong with re- 
spect to the specification of binding conditions at 
the sentence level, but offers no opportunity at all 
to extend its analytic scope beyond that sentential 
level. We claim, however, that the dependency- 
based grammar model underlying ParseTalk 
1. covers intra-sentential anaphora at the same 
level of descriptive adequacy as current GB, al- 
though it provides less complex representation 
structures than GB analyses; these structures 
are nevertheless expressive enough to capture 
the relevant distinctions, 
2. does not exhibit an increasing level of struc- 
tural complexity when faced with crucial lin- 
guistic phenomena which cause considerable 
problems for current GB theory, 
3. goes beyond GB in that it allows the treat- 
ment of anaphora at the text level of descrip- 
tion within the same grammar formalism as is 
used for sentence level anaphora, and, 
4. goes beyond the anaphora-centered treatment 
of text structure characteristic of the DRT ap- 
proach in that it already accounts for the reso- 
lution of text-level ellipsis (sometimes also re- 
ferred to as functional anaphora, cf. Hahn and 
Strube (1995)) and the interpretation of text 
macro structures (a preliminary study is pre- 
sented in Hahn (1992)). 
2 DG Constraints on Anaphora 
In this section, we present, quite informally, some 
constraints on intra-sentential anaphora in terms 
of dependency grammar (DG). We will reconsider 
these constraints in Section 3, where our grammar 
model is dealt with in more depth. We provide 
here a definition of d-binding and two constraints 
which describe the use of reflexive pronouns and 
anaphors (personal pronouns and definite noun 
phrases). These constraints cover approximately 
the same phenomena as the binding theory of GB 
(Chomsky (1981); for a computational treatment, 
cf. Correa (1988)). 
Dependency structures, by definition, refer to 
the sentence level of linguistic description only. 
The relation of dependency holds between a lexi- 
cal head and one or several modifiers of that head, 
such that the occurrence of a head allows for the 
occurrence of one or several modifiers (in some 
pre-specified linear ordering), but not vice versa. 
Speaking in terms of dependency structure rep- 
resentations, the head always precedes and, thus, 
(transitively) governs its associated modifiers in 
the dependency tree. This basic notion of govern- 
ment must be further refined for the description 
of anaphoric relations in dependency trees (we do 
not claim a universal status for the following con- 
straints, but restrict their validity to the descrip- 
tion of the German language): 
D-bindlng: A modifier M is d-bound by some 
head H, if no node N intervenes between M and H 
for which one of the following conditions holds: 
(i) node N represents a finite verb, or 
(ii) node N represents a noun with a possessive 
modifier, i.e., possessive determiners, Saxon geni- 
tive, genitival and prepositional attributes. 
Based on the definition of d-binding, we are able 
to specify several constraints on reflexive pronouns 
and anaphors in DG terms: 
Reflexive pronoun: 
The reflexive pronoun and the antecedent to which 
the reflexive pronoun refers are d-bound by the 
same head. 
Pronominal and nominal anaphors: 
(i) The antecedent c~ to which an anaphor ~r 
refers must not be governed by the same head 
which d-binds 7r, unless (ii) applies. 
(ii) The antecedent ~ to which an anaphor r refers 
may only be governed by the same head H1 which 
d-binds 7r, if (~ is a modifier of a head Hg, H2 is 
governed by H1, and a precedes 7r in the linear 
sequence of text items. Hence, a is not d-bound 
by the head H1 which d-binds 7r. I 
We will now illustrate the working of these con- 
straints, starting with the consideration of reflex- 
ives. Usually, the antecedent of a reflexive pro- 
noun is the subject of the clause to which the re- 
flexive belongs. In (1), the subject Maria is d- 
bound by the same head as the reflexive sich. 
(1) Mariai w~scht sichi. 
\[Maryi washes herselfi.\] 
Of course, the government relation between an- 
tecedent and reflexive need not be an immediate 
one. For instance, a preposition may occur be- 
tween reflexive and verb, since the notion of d- 
binding does not discriminate between NPs and 
PPs (of. (2)). If the finite verb is a modal or aux- 
iliary verb, one or more non-finite verbs may occur 
between the reflexive and the finite verb (cf. (3)). 
1 The definition of d-binding roughly corresponds to 
the governing categoryin GB terminology, which relies 
upon the notion of c-command, while the latter two 
grammar constraints correspond to the three major 
binding principles of GB. 
238 
(2) Mariai lacht fiber sichi. \[Mary/ 
laughs about herself~.\] 
(3) Maria/ m5chte sichi verbessern. 
\[Maryi wants to improve herselfi.\] 
If an intermediate node occurs between reflexive 
and antecedent which denotes a noun with a pos- 
sessive modifier, the reflexive is d-bound by this 
noun (cf. (4) vs. (5)); hence, that modifier is the 
antecedent of the reflexive. Though Maria is the 
subject of (4), only Peter can be considered the 
antecedent of the reflexive, since it is d-bound by 
the head which d-binds Peter, viz. Geschichte (cf. 
Fig. 1). If the intermediate noun has no posses- 
sive modifiers, the subject of the entire clause is 
the antecedent of the reflexive, since the reflexive 
is d-bound by the finite verb irrespective of the 
occurrence of the (object) NP (cf. (6)). 
erz ~a.l.t 
Maz£1 Gemo..h£c~te 
pObj~ 
Figure 1: Dependency Tree for Examples 4 and 5 
(4) Maria erz~.hlt Peters/ Geschichte fiber sich/. 
\[Mary tells Peter's/ story about himself/.\] 
(5) * Maria/ erz£hlt Peters Geschichte fiber sich/. 
\[* Mary/ tells Peter's story about herself/.\] 
(6) Maria/ erz~hlt eine Geschichte fiber sichi. 
\[Maryi tells a story about hersel/i.\] 
We will now consider constraints on intra- 
sentential anaphora (personal pronouns and def- 
inite NPs). As a general rule, the anaphor must 
not occupy the position of the reflexive pronoun. 
Hence, for all the examples given above, any of 
those sentences becomes ungrammatical if the re- 
flexives are replaced by non-reflexive anaphoric 
expressions (of. (7) vs. (2)). 
(7) * Maria/ lacht fiber siei. 
\[* Mary/ laughs about her/.\] 
It is also obvious that whenever the anaphor be- 
longs to a clause which is subordinate to one that 
contains the antecedent, both may be coreferent: 
this holds independently of the ordering of an- 
tecedent and pronoun (cf. (8) vs. (11)). On the 
other hand, if the anaphor belongs to the ma- 
trix clause and the antecedent to the subordinate 
clause, coreference is excluded (cf. (9)). But one 
can easily think of cases where this rule is overrid- 
den. Consider, e.g., a subordinate clause preced- 
ing its matrix, as is always true for topicalizations. 
The claim that a pronoun in the superordinate 
clause must not be coreferent to an antecedent 
in a subordinate clause is then obviously false (cf. 
(10) and (11)). In (10), the antecedent Peteris not 
d-bound by the head which d-binds the anaphor 
er, and Peter precedes er. Therefore, coreference 
is possible. 2 
(8) Peter/ erwartet, daft eri einen Brief bekommt. 
\[Peter/ expects that he/ will get a letter.\] 
(9) * Er/ erwartet, daft Peter/ e/hen Brief bekommt. 
\[ * He/ expects that Peter/ will get a letter.\] 
(10) DaB Peteri e/hen Brief bekommt, erwartet er/. 
\[That Peter/ will get a letter, hei expects.\] 
(11) DaB er/ e/hen Brief bekommt, erwartet Peter/. 
\[That hel will get a letter, Peter/ expects.\] 
Another special case arises if the antecedent is a 
modifier of the subject of a sentence. In this case 
the antecedent of a pronoun may be governed by 
the head which d-binds the pronoun. In (12) the 
pronoun belongs to the subordinate clause, but in 
(13) the antecedent of the pronoun belongs to the 
subordinate clause, and the example seems to be 
acceptable. In (14), where the subject Valet is 
modified by the genitive attribute des Gewinners, 
the antecedent is governed by the head which also 
d-binds the pronoun. Both the relative clause and 
the genitive attribute are modifiers of the subject, 
which usually occurs at the first position in the 
German main clause. In this case, the antecedent 
precedes the anaphor. Hence, coreference between 
anaphor and antecedent must be granted. 3 
(12) Der Mann, der siei kennt, grfiflt die Fraui. 
\[The man who knows heri greets the womani.\] 
(13..) Der Mann, der die Frau/ kennt, grfiflt siei. 
\[The man who knows the woman/ greets heri.\] 
(14) Der Vater des Gewinnersi gratuliert ihm/. 
\[The winner's/father congratulates himi.\] 
The incorporation of an ordering constraint is 
even more justified if one looks at sentences which 
have a similar structure, but are different with re- 
spect to word order (cf. (15) vs. (16)). In (15), the 
subordinate clause immediately follows its head 
word, while in (16) the subordinate clause is ex- 
traposed. In (16) the anaphor precedes its an- 
tecedent, which is governed by the head that d- 
binds the anaphor. This violates the given con- 
straints, hence coreference is excluded. 
2GB explains topicalizatiou with a move of the top- 
icalized CP into the SpecComp phrase of the highest 
CP, so that the pronoun does not c-commandits an- 
tecedent (in these cases movements into an A-position 
are assumed for which the binding principles of GB do 
not apply). 
3GB explains this phenomenon by linking the mod- 
ifiers to the subject as adjuncts. In this position, the 
pronoun does not c-command the antecedent, and the 
adjunct of the subject is also in an A-position. 
239 
(15) Die Frage, ob Peteri nach Dublin fahren sollte, 
konnte eri noch nicht beantworten. 
\[The question, whether Peteri should go to Dublin, 
hei couldn't decide.\] 
(16) * Die Frage konnte eri noch nicht beantworten, 
ob Peteri nach Dublin fahren sollte. 
\[* The question hei couldn't decide, whether Peteri 
should go to Dublin.\] 
Structural constraints are necessary conditions, 
but additional criteria have to be considered when 
determining the antecedent of an anaphor. Mor- 
phosyntactic conditions require that a pronoun 
must agree with its antecedent in gender, num- 
ber and person, while a definite NP must agree 
with its antecedent in number only. Moreover, 
conceptual criteria have to be met as in the case 
of nominal anaphors which must subsume their 
antecedents at the conceptual level. Similarly, 
for pronominal anaphors the selected antecedent 
must be permitted in those conceptual roles con- 
necting the pronominal anaphors and its gram- 
matical head. 
The DG constraints for the use of reflexives and 
intra-sentential anaphora cover approximately the 
same phenomena as GB, but the structures used 
by DG analysis are less complex than those of 
GB and do not require the formal machinery 
of empty categories, binding chains and complex 
movements (cf. Lappin and McCord (1990, p.205) 
for a similar argument). Hence, our proposal pro- 
vides a more tractable basis for implementation. 
3 Major Grammatical Predicates 
The ParseTalk model of DG (Hahn et al., 1994) 
exploits inheritance as a major abstraction mech- 
anism. The entire lexical system is organized as 
a hierarchy of lexical classes (isac denoting the 
subclass relation among lexical classes), with con- 
crete lexical items forming the leave nodes of the 
corresponding lexicon grammar graph. Valency 
constraints are attached to each lexical item, on 
which the local computation of concrete depen- 
dency relations between a head and its associated 
modifier is based. These constraints incorporate 
categorial knowledge about word classes and mor- 
phosyntactic knowledge involving complex feature 
terms as used in unification grammars. 
The definition of the grammatical predicates be- 
low is based on the following conventions: II de- 
notes the unification operation, 2_ the inconsis- 
tent element. Let u be a complex feature term 
and ! a feature, then the extraction u\l yields 
the value of l in u. By definition, u\l gives _l_ 
in all other cases. In addition, we supply ac- 
cess to conceptual knowledge via a KL-ONE-style 
classification-based knowledge representation lan- 
guage. The concept hierarchy consists of a set of 
concept names ~" = {COMPUTERSYSTEM, NOTE- 
BOOK, MOTHERBOARD, ...} and a subclass rela- 
tion isaF = {(NOTEBOOK, COMPUTERSYSTEM), 
(PCI-MOTHERBOARD, MOTHERBOARD), ...} C 
x 9 r. roles C f" x 9 v is the set of relations 
with role names "R = {has-part, has-cpu, ...} and 
denotes the established relations in the knowledge 
base, while R characterizes the labels of admitted 
conceptual relations. The relation permit C 9 v x 
"R x 9 r characterizes the range of possible con- 
ceptual relations among concepts, e.g., (MOTn- 
ERBOARD, has-cpu, CPU) E permit. Furthermore, 
object.attribute denotes the value of the property 
attribute at object and the symbol self refers to 
the current lexical item. The ParseTalk specifi- 
cation language, in addition, incorporates topo- 
logical primitives for relations within dependency 
trees. The relations left and head denote "~ oc- 
curs left of y" and "x is head of y', resp. These 
primitive relations can be considered declarative 
equivalents to the procedural specifications used 
in several tree-walking algorithms for anaphora 
resolution, e.g., by Hobbs (1978) or Ingria and 
Stallard (1989). Note that in the description be- 
low tel + and rel* denote the transitive and transi- 
tive/reflexive closure of a relation rel, respectively. 
x d-binds y :<:~ 
(x head + y) 
A -~q z: ((x head + z) A (z head + y) 
A (z isac* finiteVerb 
V 3u: (z head u 
A ((z spec u A u isac* DetPossessive) 
V (z saxGen u A u isac* Noun) 
V (z ppAtt u A u isac* Noun) 
V (z genAtt u A u isac* Noun))))) 
Box 1: D-binding 
The possible antecedents that can be reached 
via anaphoric relations are described by the pred- 
icates isPotentialReflexi~eAntecedentOf (eft Box 
2) and isPotentialAnaphoricAntecedentOf(cf. Box 
3). These incorporate the predicate d-binds (of. 
Box 1) which formally defines the corresponding 
notion from Section 2. The evaluation of the ma- 
x isPotentialReflexiveAntecedentOf y :¢¢. 
3 z: (z d-binds y A z d-binds x) 
Box 2: isPotentialRefleziveAntecedentOf 
jor predicate, isPotentiaIAnaphoricAntecedentOf 
(cf. Box 3), determines the candidate set of possi- 
ble antecedents for (pro)nominal anaphors, and 
240 
x isPotentialAnaphoricAntecedentOf y :~:~ 
--,3 z: (z d-binds x A z d-binds y) 
A (x left + y 
V --,=1 u: (u d-binds y 
A (u head + x))) 
Box 3: isPotentialAnaphoricAntecedentOf 
thus characterizes the notion of teachability in 
formal terms. The use of constraints as filters 
becomes evident through the further restriction 
of this set by the predicates adapted to partic- 
ular grammatical relations, thus taking the no- 
tion of satisfiability into account. For instance, 
the predicate PronAnaphorTest from Box 4 con- 
tains the grammatical constraint for pronominal 
anaphors according to which some pronoun and 
its antecedent must agree in gender, number, and 
person, and the conceptual constraint described 
in Section 2. The predicate NomAnaphorTest 
from Box 5 captures the conceptual constraint 
for nominal anaphors such that the concept to 
which the antecedent refers must be subsumed by 
the concept to which the anaphoric noun phrase. 
refers. Additionally it tests whether the defi- 
nite NP agrees with the antecedent in number. 
These two predicates cover the knowledge related 
to the resolution of intra-sentential as well as 
inter-sentential anaphora. Note the equivalence 
of grammatical and conceptual conditions within 
a single constraint. All these predicates form part 
of the computation process aiming at the resolu- 
tion of anaphora as described in Section 4. 
PronAnaphorTest (pro, ante):¢:~ 
ante isac* Noun A 
((pro.features\selt~agr\gen) 
LJ(ante.features\self~agr\gen) ¢ _L) A 
((pro.features \self\agr\num) 
U(ante.features\self~agr\nurn) # _L) A 
((pro.features\self~agr\pers) 
Ll(ante.features\self~agr\pers) # .1_) ^ 
Vx V role E R: 
(x head pro A 
(x.concept, pro.concept) E roles 
=~(x.concept, role, ante.concept) E permit) 
Box 4: PronAnaphorTest 
NomAnaphorTest (defNP, ante):~ 
ante isac* Noun A 
((defNP.features \sel~agr\num) 
U(ante.features\sel~ag,\num) # _L) ^ 
ante.concept isaF* defNP.concept 
Box 5: NomAnaphorTest 
4 Resolution of Anaphora 
The ParseTaik environment builds on the actor 
computation model (Agha and Hewitt, 1987) as 
background for the procedural interpretation of 
lexicalized dependency specifications in terms of 
so-called word actors (of. Schacht et al. 1994; 
Hahn et al. 1994). Word actors combine object- 
oriented features with concurrency yielding strict 
lexical distribution and distributed computation 
in a methodologically clean way. The model 
assumes word actors to communicate via asyn- 
chronous message passing. An actor can send mes- 
sages only to other actors it knows about, its so- 
called acquaintances. The arrival of a message at 
an actor is called an event; it triggers the execu- 
tion of a method that is composed of atomic ac- 
tions - among them the evaluation of grammatical 
predicates. As we will show, the specification of 
a particular message protocol corresponds to the 
treatment of fairly general linguistic tasks, such 
as establishing dependencies, properly arranging 
coordinations, and, of course, resolving anaphors. 
Consequently, any of these subprotocols consti- 
tutes part of the grammar specification proper. 
We shall illustrate the linguistic aspects of 
word actor-based parsing by introducing the basic 
data structures for text-level anaphora as acquain- 
tances of specific word actors, and then turn to the 
general message-passing protocol that accounts 
for intra- as well as inter-sentential anaphora. 
Our exposition builds on the well-known focus- 
ing mechanism (Sidner, 1983; Grosz and Sidner, 
1986). Accordingly, we distinguish each sentence's 
unique focus, a complementary list of alternate 
potential loci, and a history list composed of dis- 
course elements not in the list of potential loci, 
but occurring in previous sentences of the current 
discourse segment. These data structures are re- 
alized as acquaintances of sentence delimiters to 
restrict the search space beyond the sentence to 
the relevant word actors. 
The protocol level of analysis encompasses the 
procedural interpretation of the declarative con- 
straints given in Section 2. At that level, in the 
case of reflexive pronouns, the search for the an- 
tecedent is triggered by the occurrence of a reflex- 
ive pronoun in the text. Upon instantiation of the 
corresponding word actor, a SearchRefAntecedent 
message will be issued. The distribution strategy 
of the message incorporates the syntactic restric- 
tions for the appearance of a reflexive pronoun and 
its possible antecedent. This can be described in 
terms of two phases: 
1. In phase 1 the message is forwarded from its 
initiator to the word actor which d-binds the 
241 
initiator. Only if the message reaches a finite 
verb or a noun which has a possessive modifier 
is a new message with phase 2 sent, and the 
message in phase I terminates. On any other 
occasion (e.g., the head of the initiator is a 
preposition or a non-finite verb) the message 
is simply passed on to the receiver's head. 
2. In phase 2 the message is forwarded to the 
subject of the finite verb or, if a noun d-binds 
the reflexive, to the possessive modifier of the 
noun. 
Only nouns or personal pronouns are capable of 
responding to SearchRefAntecedenl messages. If 
the search for an antecedent is successful, a Ref- 
AntecedentFound message is sent directly to the 
initiator of the search message which changes its 
concept identifier accordingly. 
For pronominal anaphors, the search for the an- 
tecedent is triggered by the occurrence of a per- 
sonal pronoun. Upon instantiation of the cor- 
responding word actor, a SearchPronAnlecedent 
message will be sent. For nominal anaphors, the 
search for the antecedent is triggered by the at- 
tachment of a definite article as a modifier to its 
head noun, so that a SearchNomAntecedent mes- 
sage will be issued. Since the structural criteria 
for the sentence position of both types of anaphors 
are the same, the distribution mechanisms under- 
lying the corresponding messages can be described 
by their common superclass, SearehAntecedent. Its 
distribution strategy incorporates the syntactic 
restrictions for the appearance of both elements 
involved, anaphor and antecedent. This can be 
described in terms of three main phases: 
1. In phase 1, the message is forwarded from its 
initiator to the head which d-binds the initia- 
tor. Only if the message reaches this head are 
two further messages with phases la and 2 sent 
simultaneously, and the message in phase I ter- 
minates. On any other occasion (e.g., the head 
of the pronoun is a preposition) the message is 
simply passed on to the receiver's head. 
(a) In phase la the modifiers of the initia- 
tor's direct head are tested, in order to 
determine if any of these modifiers have 
modifiers themselves. When the test suc- 
ceeds, the message is forwarded to these 
modifiers, where the anaphor predicates 
( PronAnaphorTest or NomAnaphorTest) 
are evaluated in parallel. 
2. In phase 2 the message is forwarded from the 
head which d-binds the initiator (the original 
sender) to the word actor which represents the 
sentence delimiter of the current sentence. If on 
that path the message encounters a head which 
d-binds the sender (mediating messages from 
the initiator), that head may possibly govern 
an antecedent in its subtree. New messages 
with phase 2a are sent (their number depends 
on how many modifiers of the head exist). 
(a) In phase 2a the message is forwarded from 
the head which d-binds the sender to each 
of its modifiers (excluding the sender of the 
message), where both anaphor predicates 
are evaluated. 
3. Phase 3 is triggered independently from phase 
1 and 2. The path leads from the initiator to 
the sentence delimiter of the previous sentence, 
where its state is set to phase 3a. 
(a) In phase 3a the sentence delimiter's ac- 
quaintances Focus and PotFoci are tested 
for the anaphor predicates. 
Note that only nouns or personal pronouns are 
capable of responding to SearchAn~ecedent mes- 
sages and test whether they fulfill the required 
criteria for an anaphoric relation. If any of the 
anaphor predicates succeeds, the determined an- 
te6edent sends an AntecedentFound message di- 
rectly to the initiator of SearchAntecedent; this 
message carries the concept identifier of the an- 
tecedent. The initiator of the SearchAnteeedent 
message, viz. the anaphor, upon receipt of the An- 
tecedentFound message changes its concept identi- 
fier accordingly. This update of the concept iden- 
tifier is the final result of anaphora resolution, a 
change which accounts for the coreference between 
concepts denoted by different lexical items at the 
text level. 
We now discuss the protocol for establishing 
anaphoric relations based on intra- and inter- 
sentential anaphora considering the following text: 
(17) Die Firma Cornpaq, die den LTE-Lite entwik- 
kelt, bestiickt ihn mit einem PCI-Motherboard. 
Der Rechner hat eine Taktfrequenz yon 50 Mhz. 
\[The company Compaq, which develops the LTE- 
Lite, equips it with a PCI-motherboard. The sys- 
tem comes with a clock frequency of 50Mhz.\] 
In the first sentence of (17), the SearchAnteeedent 
message is caused by the occurrence of the per- 
sonal ihn (cf. Fig. 2 which depicts two instances 
of anaphora resolution). In phase 1, the message 
reaches the finite verb bestiickt, where two new 
instances of the message are created. In phase 
2 it takes the path to the sentence delimiter of 
the current sentence (no effect). In phase la, the 
message reaches the subject Firma, which is the 
leftmost modifier of the verb, and determines the 
noun LTE-Lite as the only possible antecedent of 
ihn. The success of PronAnaphorTesl leads to the 
sending of an AntecedentFound message, the re- 
sult of which is the update of the concept identi- 
242 
---~* searchA~tecedent message 
...... :~ antecedentFound message 
\[.1 
~iiiii!iiiiii I "o''~:','--'~-.~ ..... "~0"'1 iiiiii::::::i:i ::~ 
.... :;!iiiii L ~'~L"~' " • J iiii::i:: :.:'!i:: 
......... ~`~!~ii~i~ii~ii!iii~iiiiiiiiiii~i!iiiii~!iiiiiii~i~iiiiiii~iiiiiii~i~iiiii;i~:i~::~ ........... 
la ....... ~ i ~ • $ s t 
x I . ~ \[hat\] 
/ \ 
- \[ihn\] \[mit\] t' \[Rechne:\] . . 
\[Die\] \[Compaq\] \[entwiekelt \] .-"" • \[PCI -Motherboard\] IDeal'" 
- . ....... -- ~ °.° \[die\] \[LTE Lite\] ~'" \[elaem\] ......" 
/ ........ ~-; .............. . ........ 
\[den\] ....................................................... "'" 
Die Firma Cc~paq, 
The company Compaq, 
die den LTg-Lite entwickelt, bestfickt ihn mit einem PCI-Motherboard. Der Rechner hat ... 
whiGh the LTE-Lite develops, equips it with a PCl-motherboard. The system comes ... 
Figure 2: Sample Communication Protocol 
tier of ihn with that of LTE-Lite. Simultaneously, 
a SearchPronAntecedent message in phase 3 takes 
the path to the sentence delimiter of the previ- 
ous sentence, where it evaluates PronAnaphorTest 
with respect to its acquaintances Focus and Pot- 
Foci (no effect). 
The second sentence of (17) contains the def- 
inite noun phrase der Rechner. The search of 
an antecedent is triggered by the attachment of 
the definite article to the noun. In phase 1 the 
message reaches the finite verb hat, where new 
instances of the message are created. Phase la 
yields no positive results and the message termi- 
nates. In phase 2 the message takes the path from 
the finite verb to the sentence delimiter (no effect). 
Since there are no possible antecedents within 
the sentence, in phase 3 possible antecedents are 
checked which are stored as the acquaintances Fo- 
cus and PotFoci of the sentence delimiter for the 
previous sentence. Since Rechner subsumes LTE- 
Life at the conceptual level, NomAnaphorTest 
succeeds. An Antecedent\[bund message is cre- 
ated, which changes the concept identifier of Rech- 
her appropriately. 
5 Comparison to Related Work 
From the linguistic viewpoint, sentence anaphora, 
so far, have only been sketchily dealt with by de- 
pendency grammarians, e.g., by Hudson (1984; 
1990). The most detailed description of grammat- 
ical regularities and an associated parsing proce- 
dure has been supplied by Lappin and McCord 
(1990). It is based on the format of a slot grammar 
(SG), a slight theory variant of DG. In particular, 
they treat pronominal coreference and anaphora 
(i.e., reflexives and reciprocals). Our approach 
methodologically differs in three major aspects 
from that study: First, unlike the SG proposal, 
which is based on a second-pass algorithm operat- 
ing on fully parsed clauses to determine anaphoric 
relationships, our proposal is basically an incre- 
mental single-pass parsing model. Most impor- 
tan't, however, is that our model incorporates the 
text-level of anaphora resolution, a shortcoming of 
the original SG approach that has recently been 
removed (Lappin and Leass, 1994), but still is a 
source of lots of problems. Third, unlike our ap- 
proach, even the current SG model for anaphora 
resolution does not incorporate conceptual knowl- 
edge and global discourse structures (for reasons 
discussed by Lappin and Laess). This decision 
might nevertheless cause trouble if more conceptu- 
ally rooted text cohesion and coherence structures 
have to be accounted for (e.g., textual ellipses). 
A particular problem we have not yet solved, 
the plausible ranking of single antecedents from 
a candidate set, is dealt with in depth by Lappin 
and Laess (1994) and Hajicova et al. (1992). Both 
define salience metrics capable of ordering alter- 
native antecedents according to structural crite- 
ria, several of which can directly be attributed to 
the topological structure and topic/comment an- 
notations of the underlying dependency trees. 
243 
6 Conclusions 
We have outlined a model of anaphora resolution 
which is founded on a dependency-based gram- 
mar model. This model accounts for sentence-level 
anaphora, with constraints adapted from GB, as 
well as text-level anaphora, with concepts close 
to Grosz-Sidner-style focus models. The associ- 
ated text parser is based on the actor computation 
model. Its message passing mechanisms constitute 
the foundation for expressing specific linguistic 
protocols, e.g., that for anaphora resolution. The 
main advantage of our approach lies in the unified 
framework for sentence- and text-level anaphora, 
using a coherent grammar format, and the pro- 
vision for access to grammatical and conceptual 
knowledge without prioritizing either one of them. 
It is also a striking fact that, given the same lin- 
guistic phenomena, structural dependency config- 
urations are considerably simpler than their GB 
counterparts, though suitably expressive. 
The anaphora resolution module (for reflexives, 
intra- and inter-sentential anaphora) has been re- 
alized as part of ParseTalk, a dependency parser 
which forms part of a larger text understanding 
system for the German language, currently under 
development at our laboratory. The parser has 
been implemented in Smalltalk; the Smalltalk sys- 
tem itself, which runs on a SUN SparcStation net- 
work, has been extended by asynchronous message 
passing facilities and physical distribution mecha- 
nisms (Xu, 1993). The current lexicon contains a 
hierarchy of approximately 100 word class specifi- 
cations with nearly 3.000 lexical entries and corre- 
sponding concept descriptions from two domains 
(information technology and medicine) available 
from the LOOM knowledge representation system 
(MacGregor and Bates, 1987). 
Acknowledgments. We would like to thank our col- leagues in the CZ~Z) r group who read earlier versions 
of this paper. In particular, improvements are due to 
discussions we had with S. Schacht, N. Br6ker, P. Neu- 
haus, and M. Klenner. We also like to thank J. Alcan- 
tara (CorneU U) who kindly took the role of the native 
speaker via Internet. This work has been funded by 
LGFG Baden- Wiirttemberg (1.1.4-7631.0; M. Strube) 
and a grant from DFG (Ha 2907/1-3; U. Hahn). 
References 
Agha, G. and Hewitt, C. (1987). Concurrent program- 
ming using actors. In A.Yonezawa and M.Tokoro 
(Eds.), Object-Oriented Concurrent Programming. 
Cambridge/MA: MIT Pr., pp.37-53. 
Brfker, N., Hahn, U., and Schacht, S. (1994). Concur- 
rent lexicalized dependency parsing: the ParseTalk 
model. Proc. COLING '9~. Kyoto, Japan. Vol.1, 
pp.379-385. 
Chomsky, N. (1981). Lectures on Government and 
Binding. Dordrecht: Foris. 
Correa, N. (1988). A binding rule for government- 
binding parsing. Proc. COLING '88. Budapest, 
Hungary. Vol.1, pp.123-129. 
Grosz, B. J. and Sidner, C. L. (1986). Attention, in- 
tentions, and the structure of discourse. Computa- 
tional Linguistics, 1~. (3), 175-204. 
Hahn, U. (1992). On text coherence parsing. Proc. 
COLING '9~. Nantes, France, Vol. 1, pp.25-31. 
Hahn, U., Schacht, S., and BrSker, N. (1994). Con- 
current, object-oriented dependency parsing: the 
Parse Talk model. International Journal of Human- 
Computer Studies, 41. (1/2), 179-222. 
Hahn, U. and Strube, M. (1995). ParseTalk about 
Text-Level Ellipsis. Freiburg University, CLIF Tech- 
nical Report. 
Hajicova, E., Kubon, V., and Kubon, P. (1992). Stock 
of shared knowledge: a tool for solving pronominal 
anaphora. Proc. COLING "92. Nantes, France, Vol. 
1, pp.127-133. 
Hobbs, J. (1978). Resolving pronoun references. Lin- 
gua, 44. 311-338. 
Hobbs, J. R. (1982). Towards an understanding of co- 
herence in discourse. In Lehnert, W. and Ringle, M. 
(Eds.), Strategies \]or Natural Language Processing. 
Hillsdale/NJ: Erlbaum, pp. 223-243. 
Hudson, R. (1984). Word Grammar. Oxford: Black- 
well. 
Hudson, R. (1990). English Word Grammar. Oxford: 
Blackwell. 
Ingria, R. and Stallard, D. (1989). A computational 
mechanism for pronominal reference. Proc. ACL 
"89. Vancouver, Canada, pp.262-271. 
Kamp, H. and Reyle, U. (1993). From Discourse to 
Logic. Dordrecht: Kluwer. 
Kuno, S. (1987). Anaphora and discourse principles. 
In Nagao, M. (Ed.), Language and Artificial Intel- 
ligence. Amsterdam: North-Holland, pp.87-111. 
Lappin, S. and Leass, H. J. (1994). An algorithm 
for pronominal anaphora resolution. Computational 
Linguistics, ~0. (4), 535-561. 
Lappin, S. and McCord, M. (1990). Anaphora reso- 
lution in slot grammar. Computational Linguistics, 
16. (4), 197-212. 
MacGregor, R. and Bates, R. (1987). The LOOM 
Knowledge Representation Language. iSI Reprint 
Series, ISI/RS-87-188, Univ. of Southern California. 
Mann, W. C. and Thompson, S. A. (1988). Rhetorical 
structure theory: toward a functional theory of text 
organization. Text, & (3), 243-281. 
Pollard, C. and Sag, I. A. (1992). Anaphors in English 
and the scope of binding theory. Linguistic Inquiry, 
~3. (2) 261-303. 
Schacht, S., Hahn, U., and BrSker, N. (1994). Concur- 
rent lexicalized dependency parsing: a behavioral 
view on ParseTalk events. Proc. COLING '94. Ky- 
oto, Japan, Vol.l., pp.489-493. 
Sidner, C. L. (1983). Focusing in the comprehension 
of definite anaphora. In Brady, M. and Berwick, R. 
(Eds.), Computational Models o\] Discourse. Cam- 
bridge/MA: MIT Pr., pp.267-330. 
Xu, W. (1993). Distributed, Shared and Persistent Ob- 
jects. A Model \]or Distributed Object- Oriented Pro- 
gramming. London University, Dept. of Computer 
Science (Ph.D.Diss.). 
244 
