Never Look Back: An Alternative to Centering 
Michael Strube 
IRCS - Institute for Research in Cognitive Science 
University of Pennsylvania 
3401 Walnut Street, Suite 400A 
Philadelphia PA 19104 
S trube@linc, cis. upenn, edu 
Abstract 
I propose a model for determining the hearer's at- 
tentional state which depends solely on a list of 
salient discourse entities (S-list). The ordering 
among the elements of the S-list covers also the 
function of the backward-looking center in the cen- 
tering model. The ranking criteria for the S-list 
are based on the distinction between hearer-old and 
hearer-new discourse entities and incorporate pref- 
erences for inter- and intra-sentential anaphora. The 
model is the basis for an algorithm which operates 
incrementally, word by word. 
1 Introduction 
I propose a model for determining the heater's at- 
tentional state in understanding discourse. My pro- 
posal is inspired by the centering model (Grosz 
et al., 1983; 1995) and draws on the conclusions of 
Strube & Hahn's (1996) approach for the ranking of 
the forward-looking center list for German. Their 
approach has been proven as the point of departure 
for a new model which is valid for English as well. 
The use of the centering transitions in Brennan 
et al.'s (1987) algorithm prevents it from being ap- 
plied incrementally (cf. Kehler (1997)). In my ap- 
proach, I propose to replace the functions of the 
backward-looking center and the centering transi- 
tions by the order among the elements of the list of 
salient discourse entities (S-list). The S-list rank- 
ing criteria define a preference for hearer-old over 
hearer-new discourse entities (Prince, 1981) gener- 
alizing Strube & Hahn's (1996) approach. Because 
of these ranking criteria, I can account for the dif- 
ference in salience between definite NPs (mostly 
hearer-old) and indefinite NPs (mostly hearer-new). 
The S-list is not a local data structure associ- 
ated with individual utterances. The S-list rather 
describes the attentional state of the hearer at any 
given point in processing a discourse. The S-list is 
generated incrementally, word by word, and used 
immediately. Therefore, the S-list integrates in the 
simplest manner preferences for inter- and intra- 
sentential anaphora, making further specifications 
for processing complex sentences unnecessary. 
Section 2 describes the centering model as the 
relevant background for my proposal. In Section 3, 
I introduce my model, its only data structure, the 
S-list, and the accompanying algorithm. In Section 
4, I compare the results of my algorithm with the 
results of the centering algorithm (Brennan et al., 
1987) with and without specifications for complex 
sentences (Kameyama, 1998). 
2 A Look Back: Centering 
The centering model describes the relation between 
the focus of attention, the choices of referring ex- 
pressions, and the perceived coherence of discourse. 
The model has been motivated with evidence from 
preferences for the antecedents of pronouns (Grosz 
et al., 1983; 1995) and has been applied to pronoun 
resolution (Brennan et al. (1987), inter alia, whose 
interpretation differs from the original model). 
The centering model itself consists of two con- 
structs, the backward-looking center and the list 
of forward-looking centers, and a few rules and 
constraints. Each utterance Ui is assigned a list 
of forward-looking centers, C f (Ui), and a unique 
backward-looking center, Cb(Ui). A ranking im- 
posed on the elements of the Cf reflects the as- 
sumption that the most highly ranked element of 
C f (Ui) (the preferred center Cp(Ui)) is most likely 
to be the Cb(Ui+l). The most highly ranked el- 
ement of Cf(Ui) that is realized in Ui+x (i.e., is 
associated with an expression that has a valid inter- 
pretation in the underlying semantic representation) 
is the Cb(Ui+l). Therefore, the ranking on the Cf 
plays a crucial role in the model. Grosz et al. (1995) 
and Brennan et al. (1987) use grammatical relations 
to rank the Cf (i.e., subj -.< obj -< ...) but state that 
other factors might also play a role. 
1251 
Cb(Ui) = 
Cp(Vi) 
Cb(Ui) y£ 
Cp(t:i) 
For their centering algorithm, Brennan et al. 
(1987, henceforth BFP-algorithm) extend the notion 
of centering transition relations, which hold across 
adjacent utterances, to differentiate types of shift 
(cf. Table 1 taken from Walker et al. (1994)). 
Cb(Ui) = Cb(Ui-1) Cb(Ui) 
OR no Cb(Ui-1) Cb(Vi-1) 
CONTINUE SMOOTH-SHIFT 
RETAIN ROUGH-SHIFT 
Table 1: Transition Types 
Brennan et al. (1987) modify the second of two 
rules on center movement and realization which 
were defined by Grosz et al. (1983; 1995): 
Rule 1: If some element of Cf(Ui-1) is realized as 
a pronoun in Ui, then so is Cb(Ui). 
Rule 2" Transition states are ordered. CONTINUE is 
preferred to RETAIN is preferred to SMOOTH- 
SHIFT is preferred to ROUGH-SHIFT. 
The BFP-algorithm (cf. Walker et al. (1994)) con- 
sists of three basic steps: 
1. GENERATE possible Cb-Cfcombinations. 
2. FILTER by constraints, e.g., contra-indexing, 
sortal predicates, centering rules and con- 
straints. 
3. RANK by transition orderings. 
To illustrate this algorithm, we consider example (1) 
(Brennan et al., 1987) which has two different final 
utterances (ld) and (ld~). Utterance (ld) contains 
one pronoun, utterance (ld t) two pronouns. We look 
at the interpretation of (ld) and (ldt). After step 2, 
the algorithm has produced two readings for each 
variant which are rated by the corresponding tran- 
sitions in step 3. In (ld), the pronoun "she" is 
resolved to "her" (= Brennan) because the CON- 
TINUE transition is ranked higher than SMOOTH- 
SHIFT in the second reading. In (ld~), the pronoun 
"she" is resolved to "Friedman" because SMOOTH- 
SHIFT is preferred over ROUGH-SHIFT. 
(1) a. Brennan drives an Alfa Romeo. 
b. She drives too fast. 
c. Friedman races her on weekends. 
d. She goes to Laguna Seca. 
d.' She often beats her. 
3 An Alternative to Centering 
3.1 The Model 
The realization and the structure of my model de- 
parts significantly from the centering model: 
• The model consists of one construct with one 
operation: the list of salient discourse entities 
(S-list) with an insertion operation. 
• The S-list describes the attentional state of the 
hearer at any given point in processing a dis- 
course. 
• The S-list contains some (not necessarily all) 
discourse entities which are realized in the cur- 
rent and the previous utterance. 
• The elements of the S-list are ranked according 
to their information status. The order among 
the elements provides directly the preference 
for the interpretation of anaphoric expressions. 
In contrast to the centering model, my model does 
not need a construct which looks back; it does not 
need transitions and transition ranking criteria. In- 
stead of using the Cb to account for local coherence, 
in my model this is achieved by comparing the first 
element of the S-list with the preceding state. 
3.2 S-List Ranking 
Strube & Hahn (1996) rank the Cfaccording to the 
information status of discourse entities. I here gen- 
eralize these ranking criteria by redefining them in 
Prince's (1981; 1992) terms. I distinguish between 
three different sets of expressions, hearer-old dis- 
course entities (OLD), mediated discourse entities 
(MED), and hearer-new discourse entities (NEW). 
These sets consist of the elements of Prince's fa- 
miliarity scale (Prince, 1981, p.245). OLD con- 
sists of evoked (E) and unused (U) discourse entities 
while NEW consists of brand-new (BN) discourse 
entities. MED consists of inferrables (I), con- 
taining inferrables (I c) and anchored brand-new 
(BN A) discourse entities. These discourse entities 
are discourse-new but mediated by some hearer-oM 
discourse entity (cf. Figure 1). I do not assume any 
difference between the elements of each set with re- 
spect to their information status. E.g., evoked and 
unused discourse entities have the same information 
status because both belong to OLD. 
For an operationalization of Prince's terms, I stip- 
ulate that evoked discourse entitites are co-referring 
expressions (pronominal and nominal anaphora, 
previously mentioned proper names, relative pro- 
nouns, appositives). Unused discourse entities are 
1252 
-< 
Figure 1: S-list Ranking and Familiarity 
proper names and titles. In texts, brand-new proper 
names are usually accompanied by a relative clause 
or an appositive which relates them to the hearer's 
knowledge. The corresponding discourse entity is 
evoked only after this elaboration. Whenever these 
linguistic devices are missing, proper names are 
treated as unused I . I restrict inferrables to the par- 
ticular subset defined by Hahn et al. (1996). An- 
chored brand-new discourse entities require that the 
anchor is either evoked or unused. 
I assume the following conventions for the rank- 
ing constraints on the elements of the S-list. The 
3-tuple (x, uttx, posz) denotes a discourse entity x 
which is evoked in utterance uttz at the text posi- 
tion posz. With respect to any two discourse en- 
tities (x, uttz,posz) and (y, utty,pOSy), uttz and 
utty specifying the current utterance Ui or the pre- 
ceding utterance U/_ 1, I set up the following order- 
ing constraints on elements in the S-list (Table 2) 2 . 
For any state of the processor/hearer, the ordering 
of discourse entities in the S-list that can be derived 
from the ordering constraints (1) to (3) is denoted 
by the precedence relation --<. 
(I) If x E OLD and y E MED, then x -~ y. 
Ifx E OLD and y E NEW, then x -< y. 
lfx E MED and y E NEW, then x -< V. 
(2) If x, y E OLD, or x, v E MED, or x, y E NEW, 
then if uttx >- utt~, then x -< y, 
if uttz = utt~ and pos~ < pos~, then x -< y. 
Table 2: Ranking Constraints on the S-list 
Summarizing Table 2, I state the following pref- 
erence ranking for discourse entities in Ui and Ui-l: 
hearer-oM discourse entities in Ui, hearer-old dis- 
course entities in Ui-1, mediated discourse entities 
in Ui, mediated discourse entities in Ui-1, hearer- 
new discourse entities in Ui, hearer-new discourse 
entities in Ui-1. By making the distinction in (2) 
~For examples of brand-new proper names and their intro- 
duction cf., e.g., the "obituaries" section of the New York Times. 
2The relations >- and = indicate that the utterance containing 
x follows (>-) the utterance containing y or that x and y are 
elements of the same utterance (=). 
between discourse entities in Ui and discourse enti- 
ties in Ui-1, I am able to deal with intra-sentential 
anaphora. There is no need for further specifications 
for complex sentences. A finer grained ordering is 
achieved by ranking discourse entities within each 
of the sets according to their text position. 
3.3 The Algorithm 
Anaphora resolution is performed with a simple 
look-up in the S-list 3. The elements of the S-list are 
tested in the given order until one test succeeds. Just 
after an anaphoric expression is resolved, the S-list 
is updated. The algorithm processes a text from left 
to fight (the unit of processing is the word): 
1. If a referring expression is encountered, 
(a) if it is a pronoun, test the elements of the 
S-list in the given order until the test suc- 
ceeds4; 
(b) update S-list; the position of the referring 
expression under consideration is deter- 
mined by the S-list-ranking criteria which 
are used as an insertion algorithm. 
2. If the analysis of utterance U 5 is finished, re- 
move all discourse entities from the S-list, 
which are not realized in U. 
The analysis for example (1) is given in Table 3 6. 
I show only these steps which are of interest for the 
computation of the S-list and the pronoun resolu- 
tion. The preferences for pronouns (in bold font) 
are given by the S-list immediately above them. The 
pronoun "she" in (lb) is resolved to the first el- 
ement of the S-list. When the pronoun "her" in 
(lc) is encountered, FRIEDMAN is the first element 
of the S-list since FRIEDMAN is unused and in the 
current utterance. Because of binding restrictions, 
"her" cannot be resolved to FRIEDMAN but tO the 
second element, BRENNAN. In both (ld) and (ld ~) 
the pronoun "she" is resolved to FRIEDMAN. 
3The S-list consists of referring expressions which are spec- 
ified for text position, agreement, sortal information, and infor- 
mation status. Coordinated NPs are collected in a set. The S- 
list does not contain predicative NPs, pleonastic "'it", and any 
elements of direct speech enclosed in double quotes. 
4The test for pronominal anaphora involves checking agree- 
ment criteria, binding and sortal constraints. 
5I here define that an utterance is a sentence. 
61n the following Tables, discourse entities are represented 
by SMALLCAPS, while the corresponding surface expression 
appears on the right side of the colon. Discourse entitites are 
annotated with their information status. An "e" indicates an 
elliptical NP. 
1253 
(la) Brerman drives an Alfa Romeo 
S: \[BRENNANu: Brennan, 
ALFA ROMEOBN: Alfa Romeo\] 
(lb) She drives too fast. 
S: \[BRENNANE: she\] 
(1 c) Friedman 
S: \[FRIEDMANu: Friedman, BRENNANE: she\] 
races her on weekends. 
S: \[FRIEDMANu: Friedman, BRENNANE: her\] 
(ld) She drives to Laguna Seca. 
S: \[FRIEDMANE: she, 
LAGUNA SECAu: Laguna Seca\] 
(ld') She 
S: \[FRIEDMANE: she, BRENNANE: her\] 
often beats her. 
S: \[FRIEDMANE: she, BRENNANE: her\] 
Table 3: Analysis for (1) 
(2a) Brennan drives an Alfa Romeo 
S: \[BRENNANu: Brennan, 
ALFA ROMEOBN: Alfa Romeo\] 
(2b) She drives too fast. 
S: \[BRENNANE: she\] 
(2c) A professional driver 
S: \[BRENNANE: she, DRIVERBN: Driver\] 
races her on weekends. 
S: \[BRENNANE: her, DRIVERBN: Driver\] 
(2d) She drives to Laguna Seca. 
S: \[BRENNANE: she, 
LAGUNA SECAu: Laguna Seca\] 
(2d') She 
S: \[BRENNANE: she, DRIVERBN: Driver\] 
often beats her. 
S: \[BRENNANE: she, DRIVERE: her\] 
Table 4: Analysis for (2) 
The difference between my algorithm and the 
BFP-algorithm becomes clearer when the unused 
discourse entity "Friedman" is replaced by a brand- 
new discourse entity, e.g., "a professional driver ''7 
(cf. example (2)). In the BFP-algorithm, the rank- 
ing of the Cf-list depends on grammatical roles. 
Hence, DRIVER is ranked higher than BRENNAN in 
the Cf(2c). In (2d), the pronoun "she" is resolved 
to BRENNAN because of the preference for CON- 
TINUE over RETAIN. In (2d~), "she" is resolved to 
DRIVER because SMOOTH-SHIFT is preferred over 
ROUGH-SHIFT. In my algorithm, at the end of (2c) 
the evoked phrase "her" is ranked higher than the 
brand-new phrase "a professional driver" (cf. Ta- 
ble 4). In both (2d) and (2d ~) the pronoun "she" is 
resolved to BRENNAN. 
(2) a. Brennan drives an Alfa Romeo. 
b. She drives too fast. 
c. A professional driver races her on weekends. 
d. She goes to Laguna Seca. 
d/ She often beats her. 
Example (3) 8 illustrates how the preferences for 
intra- and inter-sentential anaphora interact with the 
information status of discourse entitites (Table 5). 
Sentence (3a) starts a new discourse segment. The 
phrase "a judge" is brand-new. "Mr. Curtis" is 
mentioned several times before in the text, Hence, 
7I owe this variant Andrew Kehler. -This example can mis- 
direct readers because the phrase "'a professional driver" is as- 
signed the "default" gender masculine. Anyway, this example 
- like the original example - seems not to be felicitous English 
and has only illustrative character. 
Sin: The New York Times. Dec. 7, 1997, p.A48 ("Shot in 
head, suspect goes free, then to college"). 
the discourse entity CURTIS is evoked and ranked 
higher than the discourse entity JUDGE. In the 
next step, the ellipsis refers to JUDGE which is 
evoked then. The nouns "request" and "prosecu- 
tors" are brand-new 9. The pronoun "he" and the 
possessive pronoun "his" are resolved to CURTIS. 
"Condition" is brand-new but anchored by the pos- 
sessive pronoun. For (3b) and (3c) I show only 
the steps immediately before the pronouns are re- 
solved. In (3b) both "Mr. Curtis" and "the judge" 
are evoked. However, "Mr. Curtis" is the left-most 
evoked phrase in this sentence and therefore the 
most preferred antecedent for the pronoun "him". 
For my experiments I restricted the length of the 
S-list to five elements. Therefore "prosecutors" in 
(3b) is not contained in the S-list. The discourse 
entity SMIRGA is introduced in (3c). It becomes 
evoked after the appositive. Hence SM1RGA is the 
most preferred antecedent for the pronoun "he". 
(3) a. A judge ordered that Mr. Curtis be released, but 
e agreed with a request from prosecutors that he 
be re-examined each year to see if his condition 
has improved. 
b. But authorities lost contact with Mr. Curtis after 
the Connecticut Supreme Court ruled in 1990 
that the judge had erred, and that prosecutors 
had no right to re-examine him. 
c. John Smirga, the assistant state's attorney in 
charge of the original case, said last week that 
he always had doubts about the psychiatric re- 
ports that said Mr. Curtis would never improve. 
9I restrict inferrables to the cases specified by Hahn et al. 
(1996). Therefore "prosecutors" is brand-new (cf. Prince 
(1992) for a discussion of the form of inferrables). 
1254 
(3a) A judge 
S: \[JUDGEBN: judge\] 
ordered that Mr. Curtis 
S: \[CURTISE: Mr. Curtis, JUDGEBN: judge\] 
be released, but e 
S: \[CURTISE: Mr. Curtis, JUDGEE: e\] 
agreed with a request 
S: \[CURTISE: Mr. Curtis, JUDGEE: e, REQUESTBN: request\] 
from prosecutors 
S: \[CURTISE: Mr. Curtis, JUDGEE: e, REQUESTBN: request, PROSECUTORSBN: prosecutors\] 
that he 
S: \[CURTISE: he, JUDGEE: e, REQUESTBN: request, PROSECUTORSBN: prosecutors\] 
be re-examined each year 
S: \[CURTISE: he, JUDGEE: ~, REQUESTBN: request, PROSECUTORSBN: prosecutors, YEARBN: year\] 
to see if his 
S: \[CURTISE: his, JUDGEE: ~, REQUESTBN: request, PROSECUTORSBN: prosecutors, YEARBN: year\] 
condition 
S: \[CURTISE: his, JUDGEE: e, CONDITIONBNA : condition, REQUESTBN: request, PROSECUTORSBN: prosec.\] 
has improved. 
S: \[CURTISE: his, JUDGEE: e, CONDITIONBNA: condition, REQUESTBN: request, PROSECUTORSBN: prosec.\] 
(3b) But authorities lost contact with Mr. Curtis after the Connecticut Supreme Court ruled in 1990 that the judge had 
erred, and that prosecutors had no right 
S: \[CURTISE: his, CS COURTu: CS Court, JUDGEE: judge, CONDITIONBNA: condition, AUTH.BN: auth.\] 
to re-examine him. 
S: \[CURTISE: him, CS COURTu: CS Court, JUDGEE: judge, CONDITIONBNA: condition, AUTH.BN: auth.\] 
(3c) John Smirga, the assistant state's attorney in charge of the original case, said last week 
S: \[SMIRGAE: attorney, CASEE: case, CURTISE: him, CS COURTu: CS Court, JUDGEE: judge \] 
that he had doubts about the psychiatric reports that said Mr. Curtis would never improve. 
S: \[SMIRGAE: he, CASEE: case, REPORTSE: reports, CURTISE: Mr. Curtis, DOUBTSBN: doubts\] 
Table 5: Analysis for (3) 
4 Some Empirical Dat:i 
In the first experiment, I compare my algorithm with 
the BFP-algorithm which was in a second experi- 
ment extended by the constraints for complex sen- 
tences as described by Kameyama (1998). 
Method. I use the following guidelines for the 
hand-simulated analysis (Walker, 1989). I do not as- 
sume any world knowledge as part of the anaphora 
resolution process. Only agreement criteria, bind- 
ing and sortal constraints are applied. I do not ac- 
count for false positives and error chains. Following 
Walker (1989), a segment is defined as a paragraph 
unless its first sentence has a pronoun in subject po- 
sition or a pronoun where none of the preceding 
sentence-internal noun phrases matches its syntactic 
features. At the beginning of a segment, anaphora 
resolution is preferentially performed within the 
same utterance. My algorithm starts with an empty 
S-list at the beginning of a segment. 
The basic unit for which the centering data struc- 
tures are generated is the utterance U. For the BFP- 
algorithm, I define U as a simple sentence, a com- 
plex sentence, or each full clause of a compound 
sentence. Kameyama's (1998) intra-sentential cen- 
tering operates at the clause level. While tensed 
clauses are defined as utterances on their own, un- 
tensed clauses are processed with the main clause, 
so that the Cf-list of the main clause contains 
the elements of the untensed embedded clause. 
Kameyama distinguishes for tensed clauses further 
between sequential and hierarchical centering. Ex- 
cept for reported speech (embedded and inaccessi- 
ble to the superordinate level), non-report comple- 
ments, and relative clauses (both embedded but ac- 
cessible to the superordinate level; less salient than 
the higher levels), all other types of tensed clauses 
build a chain of utterances on the same level. 
According to the preference for inter-sentential 
candidates in the centering model, I define the fol- 
lowing anaphora resolution strategy for the BFP- 
algorithm: (1) Test elements of Ui-1. (2) Test el- 
ements of Ui left-to-right. (3) Test elements of 
Cf(Ui-2), Cf(Ui-3) .... In my algorithm steps (1) 
and (2) fall together. (3) is performed using previ- 
ous states of the system. 
Results. The test set consisted of the beginnings 
of three short stories by Hemingway (2785 words, 
153 sentences) and three articles from the New 
York Times (4546 words, 233 sentences). The re- 
suits of my experiments are given in Table 6. The 
1255 
first row gives the number of personal and posses- 
sive pronouns. The remainder of the Table shows 
the results for the BFP-algorithm, for the BFP- 
algorithm extended by Kameyama's intra-sentential 
specifications, and for my algorithm. The overall 
error rate of each approach is given in the rows 
marked with wrong. The rows marked with wrong 
(strat.) give the numbers of errors directly produced 
by the algorithms' strategy, the rows marked with 
wrong (ambig.) the number of analyses with am- 
biguities generated by the BFP-algorithm (my ap- 
proach does not generate ambiguities). The rows 
marked with wrong (intra) give the number of er- 
rors caused by (missing) specifications for intra- 
sentential anaphora. Since my algorithm integrates 
the specifications for intra-sentential anaphora, I 
count these errors as strategic errors. The rows 
marked with wrong (chain) give the numbers of er- 
rors contained in error chains. The rows marked 
with wrong (other) give the numbers of the remain- 
ing errors (consisting of pronouns with split an- 
tecedents, errors because of segment boundaries, 
and missing specifications for event anaphora). 
Hem. NYT 
Pron. and Poss. Pron. 274 302 
BFP-Algo. 
BFP/Kam. 
My Algo. 
Correct 
Wrong 
Wrong (strat.) 
Wrong (ambig.) 
Wrong (intra) 
Wrong (chain) 
Wrong (other) 
Correct 
Wrong 
Wrong (strat.) 
Wrong (ambig.) 
Wrong (intra) 
Wrong (chain) 
Wrong (other) 
Correct 
Wrong 
Wrong (strat.) 
Wrong (chain) 
Wrong (other) 
189 231 
85 71 
14 2 
9 15 
17 13 
29 32 
16 9 
193 
81 
245 
57 
3 0 
17 8 
17 27 
29 15 
15 7 
217 
57 
275 
27 
21 12 
22 9 
14 6 
576 
420 
156 
16 
24 
30 
61 
25 
438 
138 
3 
25 
44 
44 
22 
492 
84 
33 
31 
20 
Table 6: Evaluation Results 
Interpretation. The results of my experiments 
showed not only that my algorithm performed bet- 
ter than the centering approaches but also revealed 
insight in the interaction between inter- and intra- 
sentential preferences for anaphoric antecedents. 
Kameyama's specifications reduce the complexity 
in that the Cf-lists in general are shorter after split- 
ting up a sentence into clauses. Therefore, the 
BFP-algorithm combined with her specifications 
has almost no strategic errors while the number of 
ambiguities remains constant. But this benefit is 
achieved at the expense of more errors caused by the 
intra-sentential specifications. These errors occur in 
cases like example (3), in which Kameyama's intra- 
sentential strategy makes the correct antecedent less 
salient, indicating that a clause-based approach is 
too fine-grained and that the hierarchical syntactical 
structure as assumed by Kameyama does not have a 
great impact on anaphora resolution. 
I noted, too, that the BFP-algorithm can gener- 
ate ambiguous readings for Ui when the pronoun 
in Ui does not co-specify the Cb(Ui-1). In cases, 
where the Cf(Ui-1) contains more than one possi- 
ble antecedent for the pronoun, several ambiguous 
readings with the same transitions are generated. 
An examplel°: There is no Cb(4a) because no ele- 
ment of the preceding utterance is realized in (4a). 
The pronoun "them" in (4b) co-specifies "deer" but 
the BFP-algorithm generates two readings both of 
which are marked by a RETAIN transition. 
(4) a. Jim pulled the burlap sacks off the deer 
b. and Liz looked at them. 
In general, the strength of the centering model is 
that it is possible to use the Cb(Ui-t) as the most 
preferred antecedent for a pronoun in Ui. In my 
model this effect is achieved by the preference for 
hearer-old discourse entities. Whenever this prefer- 
ence is misleading both approaches give wrong re- 
sults. Since the Cb is defined strictly local while 
hearer-old discourse entities are defined global, my 
model produces less errors. In my model the pref- 
erence is available immediately while the BFP- 
algorithm can use its preference not before the sec- 
ond utterance has been processed. The more global 
definition of hearer-old discourse entities leads also 
to shorter error chains. - However, the test set is 
too small to draw final conclusions, but at least for 
the texts analyzed the preference for hearer-old dis- 
course entities is more appropriate than the prefer- 
ence given by the BFP- algorithm. 
5 Comparison to Related Approaches 
Kameyama's (1998) version of centering also omits 
the centering transitions. But she uses the Cb and 
a ranking over simplified transitions preventing the 
incremental application of her model. 
l°In: Emest Hemingway. Up in Michigan. ln. The Com- 
plete Short Stories of Ernest Hemingway. New York: Charles 
Scribner's Sons, 1987, p.60. 
1256 
The focus model (Sidner, 1983; Suri & McCoy, 
1994) accounts for evoked discourse entities explic- 
itly because it uses the discourse focus, which is de- 
termined by a successful anaphora resolution. In- 
cremental processing is not a topic of these papers. 
Even models which use salience measures for de- 
termining the antecedents of pronoun use the con- 
cept of evoked discourse entities. Haji~ov~i et al. 
(1992) assign the highest value to an evoked dis- 
course entity. Also Lappin & Leass (1994), who 
give the subject of the current sentence the high- 
est weight, have an implicit notion of evokedness. 
The salience weight degrades from one sentence to 
another by a factor of two which implies that a re- 
peatedly mentioned discourse entity gets a higher 
weight than a brand-new subject. 
6 Conclusions 
In this paper, I proposed a model for determining 
the hearer's attentional state which is based on the 
distinction between hearer-old and hearer-new dis- 
course entities. I showed that my model, though 
it omits the backward-looking center and the cen- 
tering transitions, does not lose any of the predic- 
tive power of the centering model with respect to 
anaphora resolution. In contrast to the centering 
model, my model includes a treatment for intra- 
sentential anaphora and is sufficiently well specified 
to be applied to real texts. Its incremental character 
seems to be an answer to the question Kehler (1997) 
recently raised. Furthermore, it neither has the prob- 
lem of inconsistency Kehler mentioned with respect 
to the BFP-algorithm nor does it generate unneces- 
sary ambiguities. 
Future work will address whether the text posi- 
tion, which is the weakest grammatical concept, is 
sufficient for the order of the elements of the S-list 
at the second layer of my ranking constraints. I will 
also try to extend my model for the analysis of def- 
inite noun phrases for which it is necessary to inte- 
grate it into a more global model of discourse pro- 
cessing. 
Acknowledgments: This work has been funded 
by a post-doctoral grant from DFG (Str 545/1-1) 
and is supported by a post-doctoral fellowship 
award from IRCS. I would like to thank Nobo Ko- 
magata, Rashmi Prasad, and Matthew Stone who 
commented on earlier drafts of this paper. I am 
grateful for valuable comments by Barbara Grosz, 
Udo Hahn, Aravind Joshi, Lauri Karttunen, Andrew 
Kehler, Ellen Prince, and Bonnie Webber. 

References 
Brennan, S. E., M. W. Friedman & C. J. Pollard (1987). A cen- 
tering approach to pronouns. In Proc. of the 25 th Annual 
Meeting of the Association for Computational Linguis- 
tics; Stanford, Cal., 6-9 July 1987, pp. 155-162. 
Grosz, B. J., A. K. Joshi & S. Weinstein (1983). Providing 
a unified account of definite noun phrases in discourse. 
In Proc. of the 21 st Annual Meeting of the Association 
for Computational Linguistics; Cambridge, Mass., 15- 
17June 1983, pp. 44-50. 
Grosz, B. J., A. K. Joshi & S. Weinstein (1995). Centering: 
A framework for modeling the local coherence of dis- 
course. Computational Linguistics, 21 (2):203-225. 
Hahn, U., K. Markert & M. Strube (1996). A conceptual rea- 
soning approach to textual ellipsis. In Proc. of the 12 th 
European Conference on Artificial h~telligence (ECAI 
'96); Budapest, Hungary, 12-16 August 1996, pp. 572- 
576. Chichester: John Wiley. 
Haji~ov~i, E., V. Kubofi & P. Kubofi (1992). Stock of shared 
knowledge: A tool for solving pronominal anaphora. In 
Proc. of the 14 th h~t. Conference on Computational Lin- 
guistics; Nantes, France, 23-28 August 1992, Vol. 1, pp. 
127-133. 
Kameyama, M. (1998). Intrasentential centering: A case study. 
In M. Walker, A. Joshi & E. Prince (Eds.), Centering 
Theory in Discourse, pp. 89-112. Oxford, U.K.: Oxford 
Univ. Pr. 
Kehler, A. (1997). Current theories of centering for pronoun 
interpretation: A critical evaluation. Computational Lin- 
guistics, 23(3):467-475. 
Lappin, S. & H. J. Leass (1994). An algorithm for pronom- 
inal anaphora resolution. Computational Linguistics, 
20(4):535-56 I. 
Prince, E. E (1981). Toward a taxonomy of given-new informa- 
tion. In E Cole (Ed.), Radical Pragmatics, pp. 223-255. 
New York, N.Y.: Academic Press. 
Prince, E. E (1992). The ZPG letter: Subjects, definiteness, and 
information-status. In W. Mann & S. Thompson (Eds.), 
Discourse Description. Diverse Linguistic Analyses of a 
Fund-Raisbzg Text, pp. 295-325. Amsterdam: John Ben- 
jamins. 
Sidner, C. L. (1983). Focusing in the comprehension of definite 
anaphora. In M. Brady & R. Berwick (Eds.), Con,pu- 
tational Models of Discourse, pp. 267-330. Cambridge, 
Mass.: MIT Press. 
Strube, M. & U. Hahn (1996). Functional centering. In Proc. of 
the 34 th Annual Meeting of the Association for Compu- 
tational Linguistics; Santa Cruz, Cal., 23-28 June 1996, 
pp. 270-277. 
Suri, L. Z. & K. E McCoy (1994). RAFT/RAPR and centering: 
A comparison and discussion of problems related to pro- 
cessing complex sentences. Computational Linguistics, 
20(2):301-317. 
Walker, M. A. (1989). Evaluating discourse processing algo- 
rithms. In Proc. of the 27 th Annual Meeting of the Asso- 
ciation for Computational Linguistics; Vancouver, B.C., 
Canada, 26-29 June 1989, pp. 251-261. 
Walker, M. A., M. lida & S. Cote (1994). Japanese discourse 
and the process of centering. Computational Linguistics, 
20(2): 193-233. 
