RAFT/RAPR and Centering: A 
Comparison and Discussion of Problems 
Related to Processing Complex Sentences 
Linda Z. Suri* 
Educational Testing Service 
Kathleen F. McCoy t 
University of Delaware 
1. Introduction 
Several researchers have noted the local coherence exhibited by discourse (Sidner 1979; 
Grosz, Joshi, and Weinstein 1983; Carter 1987; etc.). A primary component of this local 
coherence is the way the local focus of the discourse shifts from one sentence to the next 
and the way this shifting is marked by linguistic choices made by the writer/speaker. 
By local focus, we refer to that concept a sentence is most centrally about within the 
discourse context in which it occurs. This is sometimes called the topic or center. 
A local focusing framework typically consists of focus-tracking algorithms and 
algorithms for suggesting referents for pronouns. Such a framework can be used in 
conjunction with an inferencing mechanism to resolve pronouns (and other anaphora) 
in a Natural Language Understanding system. The focusing framework suggests a 
referent for a pronoun, and the inferencing mechanism then confirms or rejects the 
suggested referent on the basis of semantic factors, i.e., semantics, world knowledge, etc. 
The focusing framework is useful because it only requires an inferencing mechanism 
to confirm a co-specification rather than requiring an inferencing to find the referent 
independently. 
To date, there have been two major frameworks for tracking the local focus from 
one sentence to the next and for using focus during pronoun resolution. The first 
framework, Focusing, was introduced by Sidner (1979). In this squib, we will describe 
our framework, Revised Algorithms for Focus Tracking and Revised Algorithms for 
Pronoun Resolution (RAFT/RAPR), which is based on Sidner's work. RAFT/RAPR 
can be characterized as maintaining two foci for a sentence: the subject focus and the 
current focus, which very often have distinct contents. RAFT/RAPR maintains a set of 
data structures, and uses rules (which take grammatical roles into account) for pronoun 
resolution and computing the foci. Taken together, these rules describe how focus can 
(and is most likely to) shift from one sentence to the next. Note that focus tracking 
and pronoun resolution are mutually dependent processes: focus tracking is necessary 
for pronoun resolution, and pronoun resolution, in turn, affects focus tracking. 
* Educational Testing Service, Mail Stop 10-R, Rosedale Road, Princeton, NJ 08541. E-mail: lsuri@rosedale.org. 
t Dept. of Computer and Information Sciences, 103 Smith Hall, University of Delaware, Newark, DE 19716. E-mail: mccoy@udel.edu. 
(~) 1994 Association for Computational Linguistics 
Computational Linguistics Volume 20, Number 2 
Subsequent to Sidner's work, Grosz, Joshi, and Weinstein (1983) introduced center- 
ing to account for the same phenomena addressed by Sidner's algorithm. 1 Centering 
attempted to simplify processing by keeping fewer data structures than Sidner's frame- 
work did. In particular, the centering literature claims that, rather than two foci, only 
one focus is needed, termed the backward-looking center (Cb). Pronoun resolution within 
the centering framework is largely based on an ordering of preferred focus (centering) 
moves. 
Other research on discourse (e.g., Grosz 1981; Grosz and Sidner 1986; Reichman 
1978) has studied another phenomenon, the global focus of discourse. The term global 
focus generally refers to the entity or set of entities that are relevant to or salient in the 
overall discourse; the identification of global focus typically interacts with the identifi- 
cation of discourse segments. Global focus and discourse segmentation are distinct from 
the phenomenon of local focusing that is addressed in this paper. However, we should 
point out that the centering literature has noted that centering "... is intended to oper- 
ate within a \[discourse\[ segment" (Walker 1989, p. 253). In our work on RAFT/RAPR 
we do not restrict the domain of the algorithms to within a discourse segment. 
Given that multiple frameworks for focus tracking and pronoun resolution have 
emerged, we would like to do a comparison to see how the frameworks are the same 
and how they differ. Previous assessments and comparisons of local focusing frame- 
works have relied on comparing how frameworks process a small number of con- 
structed discourses, but this kind of comparison is inadequate. Instead the question 
that must be answered is which framework performs best on naturally occurring text. 
However, such a comparison is not possible at this point because no framework has 
fully specified how to handle complex sentences (see Suri \[1993\] for the details of this 
argument). 
In light of this, we propose a comparison of RAFT/RAPR and centering along two 
lines. First, it is instructive to take a careful look at how the frameworks handle certain 
kinds of constructed discourses involving simple sentences. This comparison proves 
useful for understanding why the frameworks suggest the referents that they do. It 
is interesting to note that, while the methodologies used in RAFT/RAPR and center- 
ing are quite different from one another, the frameworks very often have the same 
preferences for pronoun resolution for text that is not discourse-initial (nor discourse- 
segment-initial) and that involves only simple sentences. Despite this similarity, we 
point out places where the two frameworks differ. A major difference between center- 
ing and RAFT/RAPR is that while RAFT/RAPR stacks old focus information, center- 
ing keeps information about the previous sentence only. We show why this is problem- 
atic for centering. We point out other differences that arise because centering keeps one 
focus and does not take the grammatical roles of pronouns and potential antecedents 
into account during pronoun resolution. This difference is evident in the examples 
discussed in this paper involving discourse-initial text, and even in an example dis- 
cussed in this paper that (we believe) does not involve a discourse segment boundary. 
Note that because the centering literature claims that centering should operate only 
within a discourse segment, and because this claim is used to explain some otherwise 
problematic cases of pronoun use, not being able to adequately handle discourse seg- 
1 Notice that we use the term focusing to cover all local focusing frameworks, Sidner's focusing 
framework (Sidner 1979), Carter's extensions to Sidner's framework (Carter 1987), the centering 
framework (Grosz, Joshi, and Weinstein 1983 and others), our framework (RAFT/RAPR), PUNDIT 
(Dahl \[1986\] and others), etc. We use uppercase ("Focusing," or "Local Focusing"), or "Sidner's 
Focusing Algorithm/Framework" to refer to Sidner's work. We use RAFT/RAPR to refer to our work. 
302 
Linda Z. Suri and Kathleen F. McCoy RAFT/RAPR and Centering 
ment initial text is much more of a problem for the centering frameworks than may 
at first be apparent. 
While the observations we make in this first comparison are intriguing, it would 
be inappropriate to assess and compare the frameworks only on the basis of a handful 
of constructed texts. However, in order to do a corpus analysis to compare focusing 
frameworks one must be able to handle many kinds of complex sentences. Thus, we 
developed a methodology for determining how people process a particular kind of 
complex sentence. 2 Our second line of comparison of frameworks involves studying 
how well each framework can be extended to account for such findings. Suri (1993) 
presented preliminary results for processing sentences of the form "SX because SY," 
where SX and SY each consist of a single clause. In this squib, we report some of those 
findings, and discuss extending RAFT/RAPR and centering in light of these findings. 
In closing, we summarize the abstract similarities and differences between centering 
and RAFT/RAPR. 
2. Our Focusing and Pronoun Resolution Algorithms 
Below, we discuss the behavior of our algorithms for simple (i.e., single-clause) sen- 
tences. (See Suri \[1993\] for a fuller discussion of our focusing and pronoun resolution 
algorithms, and a discussion of how our algorithms differ from Sidner's.) 
2.1 Data Structures 
Our algorithms maintain more focusing data structures than centering does, but each 
data structure is motivated by discourse processing needs. Below are the data struc- 
tures that RAFT/RAPR uses: 
• Current Focus (CF): the item computed to be the local focus of the 
sentence. 
• Potential Focus List (PFL): all NPs other than the CF and SF, ordered 
according to the following: direct object, indirect object, all other NPs in 
surface order within the clause. 
• Subject Focus (SF): the surface subject of the clause, except in certain 
cases as mentioned later. (The need for an SF as well as a local focus is 
discussed in Section 4.3.) 
• The Potential Subject Focus List (PSFL): all NPs other than the SF and 
CF, ordered as follows: direct object, indirect object, all other NPs in 
surface order within the clause. 3 
• CF stack, SF stack, PFL stack, PSFL stack. We stack the foci and foci lists 
after each sentence. (See Section 4.1.) 
2.2 Resolving Pronouns (in Simple Sentences) 
RAFT/RAPR resolves pronouns based on the grammatical role of the pronoun and 
the focusing data structures from the previous sentence. For a nonsubject third person 
2 Suri (1993) discusses tile issues that one faces in trying to determine how to process complex sentences, 
and argues why this methodology is better than alternative methodologies. 3 Determining whether we truly need both a PFL and PSFL would require a corpus analysis, which is 
beyond the capabilities of the current technology, as discussed in Suri (1993). In fact, the motivation for 
maintaining both a PFL and a PSFL is based on processing complex sentences (see Suri 1993). 
303 
Computational Linguistics Volume 20, Number 2 
singular pronoun, our algorithm first proposes the CF (of the last sentence) as the co- 
specifier, then the SF, then members of the PFL, and then, under preferences yet to be 
determined, the members of the CF stack, SF stack, PFL stack, and PSFL stack. 
We thus prefer the pronoun to co-refer with the last focus, then the subject focus 
of the previous sentence, then some other NP introduced in the previous sentence, 
and then elements that have been in focus or mentioned in previous sentences. Each 
attempted co-specification may be rejected by a separate inferencing component on 
the basis of semantic factors (semantics, world knowledge, etc.) or based on syntactic 
constraints. 
For a subject third person singular pronoun, we first try the SF, then the CF, then 
members of the PSFL, then the stacked elements. 
2.3 SF and CF Computation (in Simple Sentences) 
For a there-insertion sentence, the SF is the deep subject of the sentence, but for most 
simple sentence types the SF is the surface subject of the simple sentence. 4 
Our algorithms compute the CF of the current sentence based on the following 
interacting criteria: 
. 
. 
. 
4. 
Co-specification: Prefer elements that co-specify an element in a focusing 
data structure over elements just introduced. If an element being talked 
about now has been talked about before, it is more likely to be the topic 
(and thus to continue to be talked about) than something that has just 
been introduced. 
The type of realization of each element: Prefer NPs realized as pronouns 
over those realized with full NPs. A pronoun is more likely to be talked 
about in subsequent text than a full NP (Brown 1983). 
To appreciate this preference, consider the following. A pronoun carries 
less semantic information than a full NP. If a writer chose to 
communicate an element using a pronoun, he or she must have believed 
that the element was highly focused enough that the reader would not 
have difficulty interpreting the pronoun without the extra semantic 
information that would be communicated with a full NP. 
Which focusing data structure, if any, is co-specified by each NP. 
In general, we believe a writer/speaker is more likely to keep discussing 
the focus than to move the local focus to some other discourse entity. 
Thus, we prefer the CF to remain constant from one sentence to the next. 
We refer to this preference as the focus retention preference. We also believe 
a writer is more likely to switch focus to an element that was just 
mentioned in the previous sentence than to one that was discussed 
earlier. Thus, in sum, we prefer for the CF to co-specify the last CF 
rather than something on the last PFL or the last SF, and we prefer for 
the CF to co-specify something on the last PFL or the last SF rather than 
a stacked element. 
Syntax: We prefer for the CF to be a nonsubject rather than a subject, 
although the CF can be the subject. 
4 We need to explore how to compute the SF for other sentences that involve pleonastic subjects. 
304 
Linda Z. Suri and Kathleen F. McCoy RAFT/RAPR and Centering 
. 
Recall that we compute the SF based on subject NPs. We prefer for the 
CF and SF to be distinct, and for the SF to be a subject and the CF to be 
a nonsubject since we believe that it is often the case that there are two 
elements that are being focused on, and we believe that a subject 
pronoun is more likely to refer back to a subject and a nonsubject 
pronoun is more likely to refer back to a nonsubject. Thus, we want to 
record information in the focusing data structures in a manner that will 
allow us to take the grammatical roles of potential antecedents into 
account. 
Syntactic forms and clue words: The use of a particular linguistic form 
(e.g., a there-insertion structure) or clue words (e.g., "but") also 
influences the choice of the local focus. 
Again, these (abstract) criteria interact in determining the choice of the CE For ex- 
ample, among nonsubject NPs co-specifying elements of the PFL, we prefer a pronoun 
to a full NP; on the other hand, if the only pronoun is the subject, and it co-specifies 
a member of the PFL, and a nonsubject nonpronoun co-specifies the CF, we will not 
shift the CF to a member of the last PFL. That is, the CF will remain the same despite 
the pronominalization. Also, we prefer a nonsubject nonpronominal NP co-specifying 
a member of the PFL over a subject pronoun co-specifying the CF. 
2.4 An Example of Pronoun Resolution with RAFT/RAPR 
The following example (modified from Brennan, Friedman, and Pollard \[1987\]) illus- 
trates RAFT/RAPR. 
Example 1 
($1) Susan drives a Ferrari. SF= \[Susan\], CF=\[Ferrari\] (based on surface syntactic 
preferences) 
($2) She drives too fast. SF= \[Susan\], CF=\[Susan\] 
($3) Lyn races her on weekends. SF= \[Lyn\], CF=\[Susan\], PFL=(\[weekends\]) 
($4) She often beats her. 
For the first sentence, we compute the SF to be the surface subject, \[Susan\], and 
the CF to be the direct object, \[a Ferrari\]. 
For $2, we first try to resolve the subject, "She," using SF(S1); since this is not 
rejected by inferencing with semantic factors and syntactic information, the pronoun 
is resolved to refer to \[Susan\]. SF(S2)= \[Susan\], the surface subject of the sentence. Since 
the only co-specifier to something previously mentioned is \[Susan\], CF(S2)=\[Susan\] as 
well. 
For $3, we try to resolve the nonsubject pronoun "her" using CF(S2). Since this 
is not rejected by inferencing with semantic factors and syntactic information, the 
pronoun is resolved to refer to CF(S2), \[Susan\]. We compute SF(S3) to be the surface 
subject, \[Lyn\]. Since \[Susan\] ("her') is the only thing that specifies anything previously 
mentioned, CF(S3)=\[Susan\]. 
In ($4), we first try to resolve "She" using the SF and "her" using the CF. This 
interpretation is not rejected on the basis of inferencing with semantic factors or syn- 
tactic information. Thus, we get the reading "Lyn often beats Susan," which is the 
same reading that native speakers of English preferred in an informal poll. 
305 
Computational Linguistics 
Table 1 
Centering transitions. 
Cb(Un)=Cb(Un-1) Cb(Un)#Cb(Un-1) 
Cb(Un)=Cp(Un) Continue Smooth-Shift 
Cb(Un)¢Cp(Un) Retain (Rough-)Shift 
Volume 20, Number 2 
No-Cb: No element in Un realizes an element of Cf(Un-l). 
3. Centering 
In this section, we give a brief introduction to the centering framework. This frame- 
work was introduced by Grosz, Joshi, and Weinstein (1983), it has been discussed 
and/or expanded on in several other works, including Brennan, Friedman, and Pol- 
lard (1987), Walker (1989, 1993), Kameyama (1986, 1993), Walker, Iida, and Cote (1992), 
Brennan (1993), Kameyama, Passonneau, and Poesio (1993), Linson (1993), and Hoff- 
man and Turan (1993). 
3.1 Computing Centers 
For each utterance, centering computes the following (Grosz, Joshi, and Weinstein 1983; 
Brennan, Friedman, and Pollard 1987; Walker, Iida, and Cote 1992): the backward-looking 
center (Cb), which is intended to capture that item which ties the current sentence in 
with the previous sentence in the discourse, and a list of forward-looking centers (Cf), 
or elements that can potentially be the Cb of the next sentence. For English, the Cf is 
ordered or ranked by grammatical relations to the main verb (Walker 1989); the order 
is "first the subject, object, and object2, followed by other subcategorized functions, 
and finally adjuncts." (Brennan, Friedman, and Pollard 1987, p. 156) 
The first element of the Cf list is the preferred center or (Cp). As the name implies, 
the Cp is the element that is considered most likely to become the Cb of the next 
utterance. 
After the first sentence, the Cb of utterance Un, Cb(Un), is the highest ranked 
element of the Cf of the previous utterance, Cf(Un-1), that is realized in the current 
utterance. We will refer to this condition as the Cb constraint. 
3.2 Resolving Pronouns 
The centering literature has identified four kinds of transitions between sentences 
(Walker, Iida, and Cote 1992), where Un is the current utterance, and Un-1 is the 
previous utterance. The transitions are described in Table 1. s 
Centering has used the following rules. 
Rule 1 
"If some element of Cf(Un-1) is realized as a pronoun in Un, then so is Cb(Un)." 
(Brennan, Friedman, and Pollard 1987, p. 156). 
Rule 2 
Continuing is preferred over retaining, retaining is preferred over smooth-shifting, 
5 In describing these transitions, we assume the sentence has a subject. 
306 
Linda Z. Suri and Kathteen F. McCoy RAFT/RAPR and Centering 
and smooth-shifting is preferred over shifting. (Brennan, Friedman, and Pollard 1987; 
Walker, Iida, and Cote 1992). 
In resolving pronouns, the Brennan, Friedman, and Pollard (1987) (and Walker, 
~ida, and Cote \[1992\]) algorithm uses Rule 1, Rule 2, and the Cb constraint. Basically, the 
algorithm generates all possible co-specifications for pronouns in the current sentence 
by generating all possible Cb and Cf pairs for the current sentence (corresponding to 
various co-specifications), and then it filters and ranks these possible co-specifications. 
Filtering eliminates co-specifications that violate the Cb constraint or Rule 1, and co- 
specifications involving contraindexing problems. The ranking of co-specifications is 
based on the Rule 2 transition preferences. In this way, the algorithm proposes co- 
specifications for pronouns that comply with the constraints and rules of centering, 
in an order that corresponds to centering's preferences for centering moves. Notice 
that the Cb of the current sentence is computed as a side effect of doing the pronoun 
resolution. 
3.3 Example 1 with Centering 
To illustrate centering, we look again at Example 1. 
Example 1 
1) Susan drives a Ferrari. Cf=(\[Susan\], \[Ferrari\]) 
2) She drives too fast. Cb= \[Susan\], Cf=(\[Susan\]) 
3) Lyn races her on weekends. Cb= \[Susan\], Cf=(\[Lyn\], \[Susan\]) 
4) She often beats her. 
According to the Cb constraint, Cb(4)=\[Lyn\], since Lyn is the highest ordered 
element of Cf(3). Thus, if "She" were \[Lyn\] and "her" were \[Susan\], then Cb(4)=\[Lyn\], 
Cp(4)=\[Lyn\], and this would be a smooth shift. But, if "She" were \[Susan\] and "her" 
were \[Lyn\], then Cb(4)=\[Lyn\], Cp(4)=\[Susan\], and this would be a shift. Thus, centering 
prefers the reading "\[Lyn\] often beats \[Susan\]." Thus, centering and RAFT/RAPR 
agree on how to resolve these pronouns. 
4. Discussion of Some Differences 
Although centering and RAFT/RAPR seem to resolve pronouns similarly in many 
cases, there are some situations where the two algorithms differ. 
4.1 Stacking Focus Information 
A major difference between the two frameworks is that by maintaining stacks of CFs 
and SFs, RAFT/RAPR allows a writer to pronominally refer to a CF or SF of a sentence 
before the previous one. Centering, which has no counterpart to our stacks, does not. 
The stacking of focusing information for RAFT/RAPR (or focusing algorithms in 
general) is motivated by examples of text in which the referent of a pronoun is more 
than one sentence back, and for which there is no indication that there is an intervening 
discourse segment. Example 2 is such an example. 
Example 2 
With a strain, he could see a glimpse of the river to the northwest. The walls were 
Sheetrock and bare. She had picked out some artwork. He determined that the Ego 
Wall would face the desk, behind the wing chairs. (Grisham 1991, p. 139). 
307 
Computational Linguistics Volume 20, Number 2 
Note that one might argue that the sentence "She had picked out some artwork" 
constitutes a discourse segment unto itself marked by the use of the past perfect (note 
the simple past is used in other sentences). 6 However, even if this were the case (and 
that sentence was ignored in the pronoun resolution), the centering framework would 
still not appropriately resolve the "He" in the final sentence because there would 
still be an intervening sentence ("The walls were Sheetrock and bare.") that does not 
contain the antecedent of "He." We can see no reason why this (other) intervening 
sentence should be seen as a discourse segment. 
4.2 Discourse-Initial Sentences 
Recall that centering bases its pronoun decisions on the type of "centering" transitions 
that would be manifested by various interpretations. These transitions require that the 
Cb of the previous sentence be defined. As a result, centering can make no predic- 
tions about pronouns until the third sentence in a discourse (segment). Consider the 
following: 
Example 3 
la) Lyn races Susan on weekends. 
2a) She races Jack during the week. 
Because no Cb is established in la (since la is discourse-initial), centering makes 
no prediction as to whether the pronoun in 2a should refer to \[Lyn\] or \[Susan\]. Our 
pronoun resolution algorithms, predict that "she" is \[Lyn\], the SF of la. 
Of course, one way around this problem is for a likely Cb for the first sentence 
to be calculated and used in pronoun resolution. For instance, because subjecthood is 
very important in centering, one might take the subject (or Cp) of a discourse-initial (or 
discourse-segment-initial) sentence as its Cb. 7 This would allow centering to resolve 
the referent of "She" for this text correctly. However, such a decision would cause 
centering to resolve "her" in 2b below incorrectly: 
Example 4 
lb) Lyn races Susan on weekends. 
2b) Jack races her during the week. 
An informal poll suggests that native speakers prefer to interpret "her" as referring 
to \[Susan\]. The problem here is that for centering, if there is a single pronoun in a 
sentence, it becomes the Cb, and centering prefers to resolve the Cb so that it co-refers 
with the Cb of the previous sentence regardless of the grammatical role of the Cb in 
the previous sentence. Thus, if centering is to stay within its paradigm which was set 
up for processing discourse segment internal pronouns, it will resolve a pronoun in 
one of these two examples inappropriately. (Notice no semantic or world knowledge 
inference could flag the mistake.) 
6 We discuss aspectual classification and the possible import for focusing and pronoun resolution in Suri 
(1993). 
7 In fact, this is what Brennan, Friedman, and Pollard (1987) appear to do for their examples. 
308 
Linda Z. Suri and Kathleen F. McCoy RAFT/RAPR and Centering 
Our algorithms consider the grammatical role played by a pronoun and its poten- 
tial antecedents. They correctly resolve the pronouns in these examples, s 
4.3 Pronominalization and Grammatical Roles 
Brennan, Friedman, and Pollard (1987) discuss an ambiguous discourse. (A modified 
version is shown in Example 5 below.) Centering resolves the ambiguous pronoun 
("She" in 4) one way, and RAFT/RAPR resolves it another way. This is not particularly 
interesting since the discourse is ambiguous. 
Example 5 
1) Susan drives a Ferrari. 
2) She drives too fast. 
3) Lyn races her on weekends. 
4) She wins a lot of trophies. 
(RAFT/RAPR: She=\[Lyn\], centering: She=\[Susan\]) 
What is more interesting is that if "her" in 3 is replaced by a full NP, then readers 
do have a strong preference for the pronoun to be resolved with \[Lyn\]. 
Example 6 
1) Susan drives a Ferrari. 
SF=\[Susan\], CF=\[Ferrari\]; 
2) She drives too fast. 
SF=\[Susan\], CF=\[Susan\]; 
3) Lyn races Susan on weekends. 
SF=\[Lyn\], CF=\[Susan\]; 
4) She wins a lot of trophies. 
(RAFT/RAPR: "She'=\[Lyn\]; centering: "She'=\[Susan\]) 
Cf=(\[Susan\], \[Ferrari\]) 
Cb=\[Susan\], Cf=(\[Susan\]) 
Cb=\[Susan\], Cf=(\[Lyn\], \[Susan\]) 
Notice that RAFT/RAPR computes the "She" in 4 to be \[Lyn\] since it prefers to 
resolve subject pronouns to be the SF of the previous sentence. This interpretation 
agrees with native speaker preferences for this text. On the other hand, centering 
computes the Cb in 3 to be \[Susan\[ and as a result it will prefer that "She" is \[Susan\] 
in 4 (by preferring a continue transition). Thus, centering's prediction does not agree 
with native speakers' preferences for this text. 
The problem for centering may be viewed as resulting from the fact that if a 
sentence, Un, has only one pronoun, then centering (Brennan, Friedman, and Pollard 
1987) makes that element the Cb(Un) (assuming it co-specifies something in Cf(Un-1)), 
and centering prefers to resolve pronouns so that Cb(Un)=Cb(Un-1) regardless of the 
grammatical roles that Cb(Un-1) played in Un-1. 
Because our algorithms track two foci and consider the grammatical roles of a 
pronoun and its possible antecedents during pronoun resolution (by preferring the 
8 Working within centering, Kameyama (1986) proposes preferences based on grammatical roles in resolving pronouns in certain situations. However, she does not address the discourse-initial problems. 
Also, Brennan, Friedman, and Pollard (1987) claim Kameyama's proposed rule falls out of their transition preferences, but whether this is true in all contexts is not immediately evident. 
309 
Computational Linguistics Volume 20, Number 2 
SF to the CF when resolving a subject, and the CF over the SF when resolving a 
nonsubject), we prefer for "She" to be \[Lyn\], the SF of 3. 9 
One way around this example is for centering to claim that 3 actually starts a new 
discourse segment. TM However, recall that centering is intended to operate within a 
discourse segment and discourse-segment-initial processing is problematic for center- 
ing (see Section 4.2). Thus, more work would still be required to enable centering to 
handle pronouns in Example 6 and the identical discourse with the fourth sentence 
replaced by "Jack races her during the week." 
4.4 What Is the Import of the Above Analysis? 
Two observations should be made at this point. First, it is very difficult to identify 
the focus of a discourse-initial sentence, and neither framework is likely to process all 
discourse-initial text correctly. This in itself is not terribly problematic, since discourse- 
initial text is relatively rare (since there is only one discourse-initial sentence for each 
naturally occurring text), u The problem for centering is compounded, however, since 
centering is intended to work within a discourse segment. This means that all discourse- 
segment-initial text will be problematic. The proportion of discourse-segment-initial text 
is much larger. This fact is much more evident if one notes that presumably all cases 
where a stack is needed by RAFT/RAPR (see Section 4.1) are cases where a centering 
theorist would need to argue for an intervening discourse segment. 
Second, the reader should observe that we have pointed out a potential problem 
for centering by citing two examples(Examples 3 and 4) that could not both be han- 
dled by centering. This is clearly insufficient. In fact, while RAFT/RAPR does handle 
both of these examples appropriately, there is no doubt that one could come up with 
a discourse-initial example where RAFT/RAPR would have difficulty. 12 The question 
that must be addressed is which framework would correctly process the most naturally 
occurring discourse-initial texts. Likewise, in Section 4.3 we showed that RAFT/RAPR 
could appropriately handle a non-discourse-initial text, which centering could not. 
But, again, the question we would want to address is which framework would cor- 
rectly process the most naturally occurring non-discourse-initial texts. The intent of the 
above analysis was to point out differences in the abstract/underlying preferences of 
the two frameworks, and how those differences manifest themselves during pronoun 
resolution. In particular, RAFT/RAPR's abstract/underlying preference for consider- 
ing both the grammatical role of a pronoun and the grammatical roles of its potential 
antecedents is not shared by centering. 
5. Extensions for Handling Complex Sentences 
While we have shown several problematic cases for centering for discourses involving 
constructed discourses consisting of simple sentences, clearly such a comparison is 
9 Note that the tracking of two foci is important in order to correctly handle this text and Example 5 with 
sentence 4 changed to "Jack races her during the week." 
10 In fact, that "Susan" is not pronominalized in $3 does seem to signal a shift in subject focus from 
\[Susan\] to \[Lyn\]. But whether this would constitute a new discourse segment is unclear. Furthermore, 
to rely on the claim that this is a new discourse segment, centering would need to better specify how 
to identify the start of a new discourse segment, and to better specify how to handle discourse-initial 
and discourse-segment-initial text. 
11 Furthermore, it is a bit odd to speak of the focus or foci (e.g., CF and SF) of a discourse-initial sentence, 
since it is usually difficult for a reader to determine what the topic of a discourse-initial sentence is 
from the discourse-initial sentence alone. 
12 In particular, we think that there are likely to be texts that do not involve the kind of "parallel verb 
structures" of "NP1 races NP2 ..." exhibited in the examples discussed above for which centering 
would predict the correct referent, but RAFT/RAPR would not. 
310 
Linda Z. Suri and Kathleen F. McCoy RAFT/RAPR and Centering 
inadequate. A better comparison of the the two frameworks might be based on a large 
corpora analysis to determine which framework gets the correct pronoun resolution 
more often and by suggesting the fewest number of referents on average. There are 
a number of problems one faces when trying to do a corpora analysis. One problem 
is that pronoun resolution is affected not only by focusing information, but also by 
semantic factors. In order to identify which framework best captures the pronoun 
resolution preferences speakers have based solely on the type of information intended to 
be captured by focusing frameworks, we would need to control for how often semantic 
factors are affecting pronoun resolution. 
In this squib, we are concerned with another, perhaps more interesting, prob- 
lem that precludes undertaking a corpora analysis: although complex sentences are 
prevalent in written English, neither framework has explicitly specified how to han- 
dle complex sentences. 13 In fact, some previous research handled complex sentences 
in an inconsistent manner (Sidner 1979; Linson 1993). Other research made claims that 
were not accurate because it did not acknowledge the need to specify how to process 
complex sentences. In particular, centering literature (Grosz, Joshi, and Weinstein 1983) 
questioned Sidner's use of an Actor Focus using text that involved complex sentences. 
Suri (1992) argued that the reason Sidner's algorithms could not handle that text was 
not because they used two foci, but because a method for processing of complex sentences 
was not speci~ed. 
The central question is whether one should process a complex sentence as mul- 
tiple sentences or as a single sentence. For instance, should the clauses of a complex 
sentence be processed in linear order, updating the focusing (e.g., RAFT/RAPR or 
centering) data structures as if one were processing a linear sequence of simple sen- 
tences? Or, is some other order more appropriate? More specifically, 1) How do you 
resolve pronouns within a complex sentence? Do you prefer to resolve a pronoun in a 
noninitial clause with elements from the previous clause (or the matrix clause) or with 
elements from the previous sentence? 2) How do you resolve pronouns in a subsequent 
sentence? In other words, how should one update focusing (e.g., RAFT/RAPR or cen- 
tering) data structures after a complex sentence? Answers must be provided based on 
the kinds of syntactic structures (e.g., sentential verb complements, conjunctions) used 
in the sentence. Answers might also depend on other things, such as whether a verb 
is subject-control or object-control (e.g., processing "NP1 promised NP2 SX" vs. "NP1 
persuaded NP2 SX'). 
Suri and McCoy (1993) proposed a two-part methodology for determining how to 
process a given kind of complex sentence. (Suri \[1993\] explains why this methodology 
is superior to alternative methodologies.) Note that a local focusing framework (e.g., 
RAFT/RAPR or centering) is intended to capture preferences for pronoun resolution 
independent of world knowledge, semantics, and rhetorical relations. Thus, the aim of 
the first part of the methodology is to identify how people prefer to resolve pronouns 
based on previous pronominalization, focus history, and syntax. Because it is difficult 
to create texts where pronoun resolution is not affected by world knowledge, seman- 
tics, and rhetorical relations, we develop texts that are intentionally semantically slanted 
for pronoun interpretation based on world knowledge, semantics, and rhetorical re- 
lations. Several semantically slanted texts are created in order to test the effect of the 
syntactic form of interest (e.g., "SX because SY') on pronoun resolution within the 
complex sentence, and in a sentence following that kind of complex sentence. We then 
13 Suri (1992) identified the need to address this problem, and made some preliminary proposals for 
particular sentence types. Walker (1989) indicated that centering needs to handle multiple subjects, but 
did not specify how to do that. 
311 
Computational Linguistics Volume 20, Number 2 
gather judgments about the acceptability of these texts from native English speakers. 
These judgments suggest that the readers' preferences are based on syntax, pronomi- 
nalization, and focusing history. For example, judgments suggested that readers prefer 
the subject of the SX clause over the subject of the previous sentence when resolving 
a subject in the SY clause of an "SX because SY" sentence. That is, when this hypothe- 
sized preference agreed with the semantic-slanting of the text, the text was judged acceptable. But 
when this hypothesized preference was at odds with the slanting of the text, the text was judged 
awkward or ambiguous. TM We then verify such findings by analyzing the use of anaphors 
(within and following sentences of the form of interest) in naturally occurring text. 
5.1 "SX because SY" Sentences 
Based on our analysis of grammaticality judgments from native speakers of seman- 
tically slanted discourses involving a simple sentence, an "SX because SY" sentence 
(Sn), and a simple sentence, we concluded (Suri and McCoy 1993): 
1. Readers prefer to resolve Subject(SX) with Subject(Sn_l). 
2. Readers prefer to resolve Subject(SY) with Subject(SX) (over 
Subject(Sn_l)). 
3. Readers prefer to resolve Subject(Sn+l) with Subject(SX). 
Note that these preferences 15 refute an assumption sometimes made by researchers 
regarding complex sentences: that the clauses of complex sentences can be processed 
in a linear order. While our findings indicate that the pronoun resolution within $2 does 
happen linearly, the appropriate contents of the focusing data structures after processing 
$2 should be much more heavily influenced by the SX clause (and not by the SY clause, 
as the previous assumption would require). 
Given these findings, we extended RAFT/RAPR to process sentences of the form 
"SX because SY" as follows: 
. 
. 
3. 
For resolving a Subject(SX) pronoun, first propose SF(Sn-1) as the 
referent. 
For resolving a Subject(SY) pronoun, first propose Subject(SX). 
Compute the SF of a sentence of the form "SX because SY" to be 
Subject(SX). 
These extensions allow RAFT/RAPR to propose antecedents for discourses of the 
same form as the discourses judged to be acceptable more efficiently than for dis- 
courses of the same form as those judged to be awkward. In this way, the RAFT/RAPR 
processing would "match" native speakers' judgments. 
5.2 Centering and "SX because SY" Sentences 
One must question whether or not the results of our analysis are directly applicable to 
the centering framework. However, we find several problematic cases for centering in 
processing this kind of complex sentence. First, consider the approach of processing 
"SX because SY" sentences one clause at a time, linearly, where the centering informa- 
tion from SY will be used in processing $3. (Note: this kind of processing for extending 
14 See Suri and McCoy (1993) or Suri (1993) for a discussion of the methodology. 
15 Results in Caramazza, Grober, Garvey, and Yates (1977) suggest that perhaps we should consider 
whether the verb in the SX clause is an NPl-bias or NP2-bias verb in formulating these preferences. 
312 
Linda Z. Suri and Kathleen F. McCoy RAFT/RAPR and Centering 
RAFT/RAPR was rejected by our analysis.) Under this approach, for Example 7, we 
find: 
Example 7 
$1) Dodge was robbed by an ex-convict 
the other night. 
S2X) The ex-convict tied him up because 
S2Y) he wasn't cooperating. 
3b) # Then he started screaming for help. 
Cb=Dodge?; Cf=Dodge, ex-con, night. 
Cb=Dodge; Cf=ex-con, Dodge; retain? 
Cb=Dodge; Cf=Dodge; continue. 
Cb=Dodge; Cf=Dodge; continue. 
This text looks highly coherent according to centering, but it was judged to be 
awkward or ambiguous by native speakers of English in a controlled survey (judg- 
ments gathered were: 3 acceptable; 23 awkward; 5 ambiguous). One might conclude 
that the apparent problem for centering stems from violating the conclusion that the 
text should not be processed linearly. Consider altering centering so that the Cb and 
Cp of $2 is chosen from the SX clause in a manner consistent with the solution for 
RAFT/RAPR that we outlined above. Let's look at Example 7 using this modification: 
Example 7 
$1) Dodge was robbed by an ex-convict 
the other night. 
$2) The ex-convict tied him up because 
he wasn't cooperating. 
S3b) # Then he started screaming for 
help. 
Cb=Dodge?; Cf=Dodge, ex-con, night. 
Cb=Dodge; Cf=ex-con, Dodge; retain? 
Cb=Dodge; Cf=Dodge; continue. 
Centering would still predict that this text should be highly coherent, although 
the text was judged awkward or ambiguous by native speakers. 16 Furthermore, let's 
look at another example: 
Example 8 
$1) Dodge was robbed by an ex-convict 
the other night. 
$2) The ex-convict tied him up because 
he wasn't cooperating. 
S3a) Then he took all the money and ran. 
Cb=Dodge?; Cf=Dodge, ex-con, night. 
Cb=Dodge; Cf--ex-con, Dodge; retain? 
Cb=ex-con; Cf=ex-con; smooth-shift 
This text would probably also be judged acceptable/coherent by centering using 
this strategy, and it is acceptable to native speakers (judgments gathered were: 29 
acceptable; 3 awkward). What is interesting to note is that this example would probably 
be considered less acceptable than Example 7, since this example ends in a smooth-shift 
while the other ends in a continue. 
5.3 A Possible Centering Explanation 
Work by Walker (1993) suggests that sentences beginning with "Now" (which she 
assumes mark new discourse segments) are more likely to be followed by a smooth- 
shift or retain transition rather than by a continue. While Walker bases her findings 
16 Note that if one were to try to use centering for natural language generation, centering would (probably) predict that one could pronominalize the subject of $3 in Example 7 (since $3 is a continue), 
while in fact a full NP is required to make the text less awkward/ambiguous. 
313 
Computational Linguistics Volume 20, Number 2 
on a (spoken English) corpora analysis and does not indicate how she processed any 
complex sentence preceding a sentence starting with "Now," her results suggest one 
should study whether sentences starting with "Then" are more likely to be followed 
by a retain or smooth-shift (than a continue). 17 If this is found to be the case, then an 
appropriate corresponding revision of centering's Rule 2 would allow centering, like 
RAFT/RAPR, to propose antecedents for discourses of the same form as Example 8, 
i.e., discourses judged to be acceptable, more efficiently than for discourses of the same 
form as Example 7, i.e., those judged to be awkward or ambiguous. Therefore, further 
study of complex sentences and clue words such as "Then" is needed. 
5.4 Conclusions Regarding "SX because SY" Extensions 
In sum, we extended RAFT/RAPR to process sentences of the form "SX because SY," in 
a manner that reflected the judgments given by native speakers, by simply specifying 
how to resolve pronouns in this kind of sentence and how to update the focusing 
data structures after processing this kind of sentence. However, there appears to be 
no similarly straightforward way to extend the existing centering framework to reflect 
these judgments. 
6. Comparing the Frameworks for Processing Simple Sentences 
There are a number of distinctions between RAFT/RAPR and centering that should be 
examined. This discussion assumes the algorithms are being applied to single-clause 
sentences. Many of the comments generalize to processing complex sentences. 
• RAFT/RAPR maintains two foci: a subject focus and a current focus. 
Centering maintains one focus, the Cb. TM 
• RAFT/RAPR resolves nonsubject pronouns in a different manner than 
subject pronouns, while centering does not. This results in RAFT/RAPR 
and centering having different preferences for Example 6 (where 
RAFT/RAPR matches native speaker intuitions), for some 
discourse-initial texts (see Section 4.2),, and for some ambiguous 
discourses (e.g., Example 5). 
• RAFT/RAPR resolves pronouns by searching data structures in an order 
based on several factors, including the importance of grammatical roles 
and preferences for focusing movement. Centering resolves pronouns by 
generating all possible co-specifications and then filtering and ranking 
them based on a number of constraints and rules. 
• By using Rule 1 to eliminate possible co-specifiers, centering mixes the 
process of pronoun resolution with focus computation. The RAFT/RAPR 
approach resolves pronouns and then updates the focus. Furthermore, 
although RAFT/RAPR recognizes pronominalization as a signal of focus 
when computing the CF, it does not apply a rule like centering's Rule I 
17 Related to this matter, when we gave readers the same discourses discussed earlier with the "Then" 
deleted from the $3 sentences, the judgments were still supportive of our previous findings, but they 
were more divided than for the discourses including the "Then." This supports the possibility that 
"Then" may be playing a role in the pronoun resolution for the $3 sentences. See Suri (1993) for further 
discussion. 
18 However, we note the Cp is very similar to the SF in that both are computed (for English) to be the 
surface subject of the sentence (for simple sentences). In this way, one could argue that centering does 
track two loci, to the same extent that RAFT/RAPR does. 
314 
Linda Z. Suri and Kathleen F. McCoy RAFT/RAPR and Centering 
. 
2. 
3. 
4. 
without considering the grammatical roles of anaphors and focusing 
history. 
Because the abstract preferences underlying the frameworks share much in 
common, the approaches very often make the same predictions. 
RAFT/RAPR's focusing and pronoun resolution algorithms reflect the 
following underlying abstract preferences: 
We prefer for the CF to be something that is co-referential with 
an element mentioned earlier in the discourse. 
We prefer for the CF to be a pronoun rather than a full definite 
description. 
We use preferences for resolving pronouns and computing the 
CF that involve the grammatical role of the pronoun. 
We prefer for the CF to be the same as the last CF and the SF to 
be the same as the last SF. 
Centering shares the first two of these in common with us. But, centering 
prefers the local focus to stay the same and for the subject to be the local 
focus (which is stronger than preference 4). As noted earlier, we do not 
apply preference 2 without regard for the grammatical roles of the NPs 
and which focusing data structures they co-specify. Furthermore, 
centering invokes the second preference during pronoun resolution, 
while we invoke it only after resolving the pronouns. In addition, 
centering requires that the focus (Cb) realizes an element in the 
immediately preceding sentence, as opposed to merely preferring for the 
focus to be something in the immediately preceding sentence but 
allowing it instead to be co-referential with an element further back in 
the preceding text (as RAFT/RAPR does). 
In sum, the order in which these abstract preferences are applied differs 
between the two frameworks. Furthermore, as a result of (some) 
differences between the frameworks with respect to these preferences, 
resolvhlg a pronoun to have the same grammatical role as its antecedent 
is less important in centering than it is in RAFT/RAPR (Section 4.3). 
RAFT/RAPR presents possible referents for pronouns one possibility at a 
time, and if pragmatic, semantic, and general knowledge inferencing 
rejects a referent, RAFT/RAPR proposes an alternative. Centering, on the 
other hand, (in addition to requiring this same kind of inferencing) 
sometimes suggests multiple possibilities for the co-specifications of 
pronouns in a sentence (Walker 1989). Is pragmatic inferencing applied 
at this point to pick among the possible antecedents? We assume this to 
be the case, otherwise it is not clear how centering will choose among 
the multiple potential referents. Yet, assuming this is the case, centering 
seems to involve more complex inferencing than our approach involves; 
inferencing for centering must pick which antecedent is better. 
RAFT/RAPR only asks inferencing to confirm a co-specification. 19 
19 We do not check for ambiguity of pronouns in the same way that Sidner did. However, even if we 
were to incorporate similar checks for ambiguity, such checks would involve using inferencing to confirm two possible co-specifications for a pronoun, while centering might require the use of 
inferencing to confirm more than two. Thus, the inferencing required by RAFT/RAPR would still be 
more limited than that required by centering. 
315 
Computational Linguistics Volume 20, Number 2 
7. Conclusions 
RAFT/RAPR and centering often make the same predictions for resolving pronouns. 
One reason for this is that the (abstract) preferences underlying the two frameworks 
share much in common. A major difference between the two approaches is that 
RAFT/RAPR keeps two foci, while centering keeps only one. In addition, RAFT/RAPR 
has more of a commitment to resolving pronouns according to grammatical role. 
In Section 5, we argued that one is able to extend RAFT/RAPR to process sentences 
of the form "SX because SY," by specifying how to resolve pronouns and how to update 
RAFT/RAPR data structures, to match speaker preferences for discourses involving 
sentences of that form. On the other hand, centering could not be extended in a similar 
manner. Recent work (Section 5.3) suggests that centering might be able to match 
judgments by modifying Rule 2; however, the appropriateness of such a modification 
requires further study. 
We presented examples of discourse-initial texts (Section 4.2) and a non-discourse- 
initial text (Section 4.3) involving only simple sentences that were problematic for 
centering. A second problem for centering results from not stacking focus information. 
Although the claims in this paper are based on a limited number of discourses, 
we noted that it is difficult to perform a corpora analysis to determine which frame- 
work performs better on average because of the prevalence of complex sentences and 
the lack of work on complex sentences in either framework. Furthermore, determining 
how to process particular kinds of complex sentences is a crucial step toward enabling 
such a corpora analysis. We discussed the results of an analysis of one kind of complex 
sentence ("SX because SY'). On the basis of this analysis, we extended RAFT/RAPR 
to handle this kind of complex sentence. In addition, we showed how the judgments 
collected from native speakers of English as part of this analysis were difficult to ex- 
plain within the current centering framework. Furthermore, the findings discussed in 
this paper indicate that one cannot evaluate the RAFT/RAPR and centering frame- 
works by simply applying the algorithms to finite clauses of naturally occurring text 
in linear order, without taking into account the occurrence of complex sentences and 
clue words. 
Acknowledgments 
We thank Sandee Carberry, John A. Hughes, 
and Jeff Lidz for their comments and 
advice. We are also very grateful to the 
many people, too numerous to name here, 
who provided us with judgments. This 
research was supported in part by NSF 
Grant #IRI-9010112. 
References 
Brennan, Susan E. (1993). "Reference and 
local structure in interactive discourse." 
In Proceedings, Workshop on Centering 
Theory in Naturally Occurring Discourse, 
Institute for Research in Cognitive 
Science, University of Pennsylvania. 
Brennan, Susan E.; Friedman, Marilyn W.; 
and Pollard, Carl J. (1987). "A centering 
approach to pronouns." In Proceedings, 
25th Annual Meeting of the Association for 
Computational Linguistics, 155-162. 
Brown, Cheryl (1983). "Topic continuity in 
written English narrative." In Topic 
Continuity in Discourse, edited by 
T. Givon, 317-341. J. Benjamins. 
Caramazza, Alfonso; Grober, Ellen; Garvey, 
Catherine; and Yates, Jack (1977). 
"Comprehension of anaphoric pronouns." 
Journal of Verbal Learning and Verbal 
Behavior, 16, 601-609. 
Carter, David (1987). Interpreting Anaphors in 
Natural Language Texts. John Wiley and 
Sons. 
Dahl, Deborah A. (1986). "Focusing and 
reference resolution in PUNDIT." In 
Proceedings, 1986 National Conference on 
Artificial Intelligence, Philadelphia, PA, 
1083-1088. 
Grisham, John (1991). The Firm. Dell. 
Grosz, Barbara J. (1981). "Focusing and 
description in natural language 
dialogues." In Elements of Discourse 
Understanding, edited by A. Joshi, 
316 
Linda Z. Suri and Kathleen F. McCoy RAFT/RAPR and Centering 
B. Webber, and I. Sag, 85-105. Cambridge 
University Press. 
Grosz, Barbara J., and Sidner, Candace L. 
(1986). "Attention, intentions, and the 
structure of discourse." Computational 
Linguistics, 12(3), 175-204. 
Grosz, Barbara J.; Joshi, Aravind K.; and 
Weinstein, Scott (1983). "Providing a 
unified account of definite noun phrases 
in discourse." Ln Proceedings, 21st Annual 
Meeting of the Association for Computational 
Linguistics, Cambridge, MA, 44-50. 
Hoffman, Beryl, and Turan, Omit Deniz 
(1993). "Word order and centering in 
Turkish." In Proceedings, Workshop on 
Centering Theory in Naturally Occurring 
Discourse, Institute for Research in 
Cognitive Science, University of 
Pennsylvania. 
Kameyama, Megumi (1986). "A 
property-sharing constraint in centering." 
In Proceedings, 24th Annual Meeting of the 
Association for Computational Linguistics, 
Cambridge, MA, 200-206. 
Kameyama, Megumi (1993). "Intrasentential 
centering." In Proceedings, Workshop on 
Centering Theory in Naturally Occurring 
Discourse, Institute for Research in 
Cognitive Science, University of 
Pennsylvania. 
Kameyama, Megumi; Passonneau, Rebecca; 
and Poesio, Massimo (1993). "Temporal 
centering." In Proceedings, 31st Annual 
Meeting of the Association for Computational 
Linguistics, 70-77. 
Linson, Brian (1993). "A distributional 
analysis of discourse transitions." In 
Proceedings, Workshop on Centering Theory in 
Naturally Occurring Discourse, Institute for 
Research in Cognitive Science, University 
of Pennsylvania. 
Reichman, Rachel (1978). "Conversational 
coherency." Cognitive Science, 2, 283-327. 
Sidner, Candace L. (1979). Towards a 
computational theory off definite anaphora 
comprehension in english discourse. Doctoral 
dissertation, MIT. 
Suri, Linda Z. (1992). "Correcting illegal NP 
omissions using local focus." In 
Proceedings, 30th Annual Meeting of the 
Association for Computational Linguistics 
(Student Session), 273-275. 
Suri, Linda Z. (1993). Extending focusing 
frameworks to process complex sentences and 
to correct the written English of proJi'cient 
signers of American Sign Language. Doctoral 
dissertation, University of Delaware. 
Suri, Linda Z., and McCoy, Kathleen F. 
(1993). "Focusing and Pronoun Resolution 
in Particular Kinds of Complex 
Sentences." In Proceedings, Workshop on 
Centering Theory in Naturally Occurring 
Discourse, Institute for Research in 
Cognitive Science, University of 
Pennsylvania. 
Walker, Marilyn A. (1989). "Evaluating 
discourse processing algorithms." In 
Proceedings, 27th Annual Meeting of the 
Association for Computational Linguistics, 
251-261. 
Walker, Marilyn A. (1993). "Initial contexts 
and shifting centers." In Proceedings, 
Workshop on Centering Theory in Naturally 
Occurring Discourse, Institute for Research 
in Cognitive Science, University of 
Pennsylvania. 
Walker, M.; Iida, M.; and Cote, S. (1992). 
"Japanese discourse and the process of 
centering." Technical Report 92-14, 
Institute for Research in Cognitive 
Science, The University of Pennsylvania. 
317 
