Proceedings of the 40th Annual Meeting of the Association for

Computational Linguistics (ACL), Philadelphia, July 2002, pp. 88-95.

Pronominalization in Generated Discourse and Dialogue

Charles B. Callaway James C. Lester

Istituto per la Ricerca Scientifica e Department of Computer Science
Tecnologica (ITC-irst), Italy North Carolina State University, USA
[email protected] [email protected]

Abstract anaphoric pronouns as described in Focus/Centering

Theory (Webber, 1979; Sidner, 1983; Grosz and Sid-
Previous approaches to pronominalization ner, 1986; Walker, 1998), ignoring the multitude of
have largely been theoretical rather than other possible types. However, it is certainly true
applied in nature. Frequently, such meth- that authors make use of pronouns which are not mo-
ods are based on Centering Theory, which tivated by anaphoric reference.
deals with the resolution of anaphoric pro-
In addition, because such approaches are oriented
nouns. But it is not clear that complex the-
towards anaphora resolution during parsing, they ig-
oretical mechanisms, while having satis-
nore structures such as the discourse plan which are
fying explanatory power, are necessary for
present during generation but not parsing. A typi-
the actual generation of pronouns. We first
cal discourse plan can include vital information for
illustrate examples of pronouns from vari-
pronominalization such as time and clause bound-
ous domains, describe a simple method for
aries, ordering of propositions, and semantic de-
generating pronouns in an implemented
tails verbal arguments. Current approaches based
multi-page generation system, and present
on Centering algorithms thus attempt to recreate a
an evaluation of its performance.
text coherence structure that duplicates work already
done by the discourse planner.
1 Introduction Finally, there are significant obstacles to verifying
Pronominalization is an important element in the au- the correctness of existing pronominalization algo-
tomatic creation of multi-paragraph and multi-page rithms for any pronominalization theory (Not, 1996;
texts using natural language generation (NLG). Au- Yeh and Mellish, 1997; McCoy and Strube, 1999;
thors routinely use pronouns in texts spanning all Henschel et al., 2000; Kibble and Power, 2000): the
types of genres, such as newspaper copy, science lack of natural language generation systems that can
fiction and even academic papers. Indeed, without produce large enough texts to bring discourse-level
pronouns, texts quickly become confusing as readers processes into play. Because of this, researchers are
begin to pay more attention to the writing style than forced to simulate by hand how their algorithms will
to the content that makes the text informative or en- work on a given text. It is also not sufficient to use
joyable (Callaway and Lester, 2001a). Even worse, template generation systems to perform this task be-
incorrect pronouns can lead readers to misinterpret cause they lack the low-level discourse representa-
the text or draw unsound inferences. tion needed to provide the information upon which
Furthermore, current pronominalization strategies most algorithms base their decisions.
are ill-equipped to deal with the wide variety of In this paper we first summarize related work
reasons that pronouns are used in naturally occur- in both anaphora resolution and anaphora genera-
ring texts. Almost without exception, they focus on tion. We next describe the range of pronoun types
that we found in a wide variety of texts. We pro- needed to determine appropriate pronominal forms.
ceed to describe an algorithm for determining ap- Second, Centering Theory is still vague on the ex-
propriate pronominalizations that uses existing NLG act definition of terms such as “segment” (Poesio et
structures and simple numeric techniques. We also al., 1999a), making it difficult to create a mutually
briefly describe an implemented generation system agreeable implementation.
that contains enough low-level discourse informa- An additional area of NLG research that deals
tion to motivate pronominalization decisions using with pronouns is that of referring expression gen-
this method. Finally, we quantitatively demonstrate eration (Appelt, 1985; Heeman and Hirst, 1986;
the performance of this simple numerical approach Claassen, 1992; Dale, 1992), which attempts to find
in both a newspaper and fictional narrative domain. the optimal noun phrase (whether full description,
definite description, deixis, pronoun, or reduced
2 Background and Related Work noun phrase) to enable a reader to mentally select the
intended referent from the set of possible referents
Because most NLG systems have focused on lin- (Reiter and Dale, 1997). Comparatively, referring
guistic phenomena at the paragraph level and be- expression generation is a process for local disam-
low, there has been intensive investigation into the biguation and is not generally concerned with single
core areas of generation that are required to pro- phenomena spanning multiple paragraphs. Because
duce them: discourse planning, sentence planning of this, and because the domains and genres we have
and surface realization. Since pronouns are more studied typically do not involve sets of very simi-
likely to be a multiparagraph, discourse-level phe- lar referents, we concentrate on discourse-motivated
nomenon, it has been possible to ignore their inclu- sources of pronominalization.
sion into working NLG systems which are not called
upon to generate lengthy passages. 3 Examples of Pronominalization
Indeed, most work on pronouns in computational Pronominalization is the appropriate determination,
linguistics has come under the heading of anaphora marking and grammatical agreement of pronouns
resolution as an element of parsing rather than the (he, she, their, herself, it, mine, those, each other,
heading of pronominalization as an element of gen- one, etc.) as a short-hand reference to an entity or
eration. Since discourse anaphora resolution was event mentioned in the discourse. As with anaphora
first studied theoretically (Grosz, 1977; Webber, resolution, the task of a pronominalization algorithm
1979; Sidner, 1983; Grosz and Sidner, 1986), it has is to correctly predict which pronoun a person would
come to be dominated by Centering Theory (Grosz prefer in the same situation. The range of possibili-
et al., 1995; Di Eugenio, 1998; Walker, 1998) which ties includes leaving the noun phrase as it is, reduc-
proposes rules for the determination of focus and ing it by removing some of its modifiers, or replac-
salience within a given segment of discourse. Rel- ing it with a pronoun construction.
atively little work has been done on alternate ap- Our corpora analyses have identified a number of
proaches to pronoun resolution (Hobbs, 1976; Bald- motivations for converting nouns into pronouns:
win, 1995).
While many NLG researchers have attempted to 1. Anaphoric pronouns: These are the most-
transfer the ideas of Centering Theory to genera- studied cases of pronoun occurrences, which
tion (Not, 1996; Yeh and Mellish, 1997; McCoy sequentially follow a specific entity known as
and Strube, 1999; Henschel et al., 2000; Kibble the referent. Anaphors are divided into two
and Power, 2000), there has yet been no substan- classes, short-distance (within the same sen-
tial return contribution to the field of anaphora res- tence) and long-distance (previous sentences).
olution. There are two principal reasons for this. But Johni had never been to New Orleans,
First, it is extremely difficult to create an NLG sys- and hei couldn’t remember if anyone in hisi
tem that generates the large quantity of texts needed family had either.
to exhibit discourse-level phenomena while consis-
tently employing the deep linguistic representations 2. Cataphoric pronouns: According to Quirk et
 al. (1985), cataphors are those pronouns which beginning of the entire text. Antecedent ref-
occur before their referents in the linear flow of erences that break these boundaries are some-
text within the same sentence, where the pro- times marked by the authors in academic texts:
noun is either at a lower structural level or is As we saw in the previous section, . . .
part of a fronted circumstantial clause or prepo-
sitional phrase which could have appeared after 7. Restrictions from modifiers: Because pronouns
the reference. Additionally, this category could cannot have modifiers like nouns, adding an ad-
include clefting pronouns. jective, relative clause, or some other modifier
Before hei joined the navy, Geraldi made prevents a noun from being replaced by a pro-
peace with his family. noun. For instance:
The mayor had already read the full proposal.
3. Pronouns Lacking Textual Antecedents: This
category includes document deixis (via a * The mayor had already read the full it.
demonstrative pronoun), authorial or reader 8. Focused nouns: Especially after a vocally
reference, and situational pronouns. stressed discourse marker (Wolters and Byron,
This is the first document to show . . . 2000) or some other marked shift in topic, a
We discuss these strategies in the next section. word that normally would be pronominalized
is often not, as in this example:
The group had never seen anything like it.
. . . and you frequently find that mice occupy
4. Reflexive and Reciprocal Pronouns: Most an important part of the modern medical labo-
verbs use special pronouns when the subject ratory. In other words, mice are especially nec-
and object corefer. A discourse history algo- essary for diagnosing human cancers . . .
rithm can employ that knowledge to mark re-
flexive and reciprocal pronouns appropriately. 9. Semantic and syntactic considerations: A
small number of semantic relations and syntac-
Kittensi often watch themselvesi in mirrors.
tic constructions prohibit pronominalization:
Baby lionsj tackle each otherj when playing.
* The stranger was just called him. (Bob)
5. Partitive pronouns: It is important to know con- * Roberta was no longer a her. (child)
ceptually what it is that the pronoun is trying * The father, a tyrant of a him, . . . (man)
to replace. Otherwise, it becomes impossible
to achieve the types of pronominalizations that 10. Optional pronominalization: Often there are
authors are routinely capable of creating. This borderline cases where some authors will use
requires accurate information in the knowledge pronouns while others won’t. A single algo-
base or linguistic structure from which the sen- rithm may be tuned to match a particular au-
tences are derived. thor’s style, but parameterization will be nec-
essary to match a variety of styles. Thus it is
As the horses ran by, she roped one. extremely difficult to exactly match any partic-
* As the horses ran by, she roped it. ular text without having the ability to adjust the
* As the horses ran by, she roped them. pronominalization algorithm.
In addition to these motivations, we identified
Pronominalization occurs equally as often in ex-
several factors that prevent pronouns from occurring
position as in dialogue, but dialogue can have
where they otherwise might:
slightly different pronominalizations depending on
6. Pronouns across boundaries: After a chapter, the relationship between the utterer and the hearer:
section or other obvious boundary, such as a
change in time, place, or both, as in (McCoy 11. Speaker self-reference:
and Strube, 1999), authors will typically “re- “John thinks John will go find John’s shoes,”
set” pronominalization just as if it were the John said.
 changes to first person singular pronouns: leaf nodes represent individual sentential elements
“I think I will go find my shoes,” John said. that are organized semantically. In addition, the ele-
ments of the discourse tree are typically rooted in the
12. Speaker references hearer(s): semantic knowledge base which the discourse plan-
“Mary should go find Mary’s shoes,” John ner drew from when constructing the discourse plan.
said. The discourse plan supplies the following informa-
tion that is useful for pronominalization:
changes to second person pronouns:
“You should go find your shoes,” John said.  Linearization: The sequencing information
stored in the discourse tree can be used to mo-
13. Reference to speaker and hearer (or to speaker tivate anaphoric and cataphoric pronouns as
and a third party): shown in items 1 & 2 of Section 3.
“John and Mary should go find John and
 Semantic Structure: The original subgraphs
Mary’s shoes,” John said.
(or semantic subnetworks) derived from the
changes to first person plural pronouns: knowledge base can motivate content vs. sit-
“We should go find our shoes,” John said. uational knowledge (item 3) reflexive and re-
ciprocal pronouns via argument lists (item 4),
14. Reference to a third party: partitive pronouns (item 5), and the existence
“Bob and Mary went to eat Bob and Mary’s of NP modifiers (item 7), and can identify se-
breakfast,” John said. mantic types in relations (item 9).
changes to third person plural pronouns:  Discourse Structure: The rhetorical relations
“They went to eat their breakfast,” John said. that hold between different sentences typically
imply where section boundaries are located
15. Finally, the treatment of pronouns differs de- (item 6), indicate what types of discourse mark-
pending if they are inside or outside of the di- ers are employed (item 8), and in the case of
rect quotation. For example: dialogue, know which actors are speaking, lis-
“Oh man, I forgot to eat my breakfast!” John tening, or not present (items 11-15).
muttered to himself while grabbing his shoes.
This detailed knowledge of the discourse is avail-
Although this enumeration is surely incomplete, able to an implemented pronominalization compo-
it provides a basic description of the types of phe- nent utilizing any theory, including Centering the-
nomena that must be handled by a generation system ory. We thus now turn our attention to what role this
in order to produce text with the types of pronouns information plays in a pronominalization algorithm.
found in routine human-produced prose.
5 A Simple Pronominalization Algorithm
4 Architectural Concerns
At an abstract level, the pronominalization algo-
In order to correctly account for these phenomena rithms derived from Centering theory are easily ex-
during generation, it is necessary to have detailed pressed: if Centering theory predicts a pronoun
information about the underlying discourse struc- would be used in anaphora resolution in a given seg-
ture. Although a template generation system could ment of text, then generate the appropriate pronoun.
be augmented to record this information, in practice While this works for many cases of anaphoric pro-
only deep structure, full-scale NLG systems have the nouns [84.7% in (McCoy and Strube, 1999), 87-
requisite flexibility. Because a pronominalization al- 90% in (Henschel et al., 2000)], we have seen that
gorithm typically follows the discourse planner, it these form only a subset of the potential reasons for
frequently has access to the full discourse plan. pronominalization. Furthermore, this approach as-
A typical discourse plan is a tree structure, where sumes that the discourse tree was constructed with
internal nodes represent structuring relations while Centering theory in mind.
Given: course tree allows for the determination of
LNE , the linearized list of nominal elements
NE , the current nominal element boundaries and thus the concept of metric dis-
SEEN , the list of encountered nominal elements tance between elements.
D, the dialogue state of the current leaf node
RS , the rhetorical structure near the leaf node  Rhetorical relations: The rhetorical relations
SC , the sentence counter
Do: can tell us which nominal elements follow dis-
SEEN ( ; SC ( 0 course markers and which are used reflexively
while LNE 6=  do
NE ( first(LNE ) or reciprocally.
if NE 62 SEEN
then reset-counters(NE ),  Dialogue: By recording the participants in dia-
SEEN ( SEEN  NE logue, the discourse tree allows for the appro-
else update-counters(NE )
D ( updateDialogueState()
priate assignment of pronouns both inside and
RS ( updateLocalRhetoricalStructure() outside of the direct quote itself.
if (topic-shift _ time-shift) 2 RS
then SC ( SC + 10 The algorithm we developed considers the cur-
else if modifiers(NE; RS ) =  ^
(special-relation _ appositive) 62 RS
rent discourse leaf node and the rhetorical structure
if D == QuotedDialogue above it, and also makes use of the following data:
then mark(quoted-pronoun(NE; RS ))
else if subject-matches-object(NE; RS )  Nominal element distance: How many total
then mark(ReflexiveP ronoun)
else if sent-distance(NE; SC ) = 0
(non-distinct) nominal elements ago a particu-
then mark(MultipleInSentenceP ronoun) lar element was last used.
else if 3 <= sent-distance(NE; SC ) < 1
and nominal-distance(NE ) < 3  Recency: How many distinct nominal elements
then mark(LongDistanceP ronoun),
else if recency (NE ) > 3
have been seen since its last use.
then mark(ShortDistanceP ronoun),
LNE ( remove-first(LNE ); SC ( SC + 1  Sentential distance: How many sentences (pro-
totypical clauses) have appeared since the last
Figure 1: The Pronominalization Algorithm usage of this nominal element.

The algorithm itself (Figure 1) is best character-

However, it is not clear that Centering theory itself ized as a counting method, that is, it loops once
is necessary in generation, let alone its accompany- through the linearized list of nominal elements and
ing algorithms and data structures. Because Cen- makes pronominalization decisions based on the lo-
tering theory is typically applied to parsing (which cal information described above, and then updates
starts with no discourse tree), it may not be the most those numerical counters. Numerical parameters
efficient technique to use in generation (which has a (e.g., recency (NE ) > 3) are derived from empir-
complete discourse tree available for inference). ical experimentation in generating multi-page prose
Instead, we attempted to determine if the informa- in a narrative domain.
tion already present in the discourse tree was enough While it lacks the explanatory power of a rela-
to motivate a simpler algorithm based on the follow- tively mature linguistic theory, it also lacks the ac-
ing available data: companying complexity and is immediately appli-
cable to real-world deep generation systems. The al-
 Ordered sequence of nominal elements: Be- gorithm is traced in Figure 2, although due to space
cause the discourse tree is linearized and in- limitations some phenomena such as dialogue, long
dividual leaves of the tree annotate which ele- distance and reflexive pronouns are not shown.
ments have certain semantic roles, a very good
guess can be made as to which nominal ele- 6 Implementation and Evaluation
ments precede others at the clause level.
S TORY B OOK (Callaway and Lester, 2001b; Call-
 Known paragraph and sentence boundaries: away and Lester, in press) is an implemented nar-
Analysis of the rhetorical structure of the dis- rative generation system that converts a pre-existing
 Sentences as seen by the reader (antecedents underlined, pronouns in bold):
Now, it happened that a wolf 1 , a very cruel, greedy creature 2 also heard Little Red Riding Hood 3 as
she 4 passed, and he 5 longed to eat her 6 for his 7 breakfast8 . But he 9 knew Hugh 10 , the woodman 11 ,
was at work12 very near with his 13 great dog14 .
Sentences as produced by the discourse planner before revision:
S1: Now, it happened that a wolf 1 , a very cruel, greedy creature 2 also heard Little Red Riding Hood 3
as Little Red Riding Hood 4 passed.
S2: The wolf 5 longed to eat Little Red Riding Hood 6 for the wolf’s 7 breakfast 8 .
S3: But the wolf 9 knew Hugh 10 , the woodman 11 , was at work 12 very near with Hugh’s 13 great dog 14 .
Each noun element is processed in the order linearized from the discourse plan:
1. The first mention of wolf1 in the narrative resets its discourse history entry.
2. Creature2 is the second mention of wolf, but it is in an appositive structure (see pronoun category #9).
3. LRRH3 was mentioned just before in the prior paragraph, but “Now,” is a prosodic discourse marker
(see pronoun category #8), thus modifiers(NE, RS) 6= .
4. For LRRH3 and LRRH4 , sentence-distance(NE, SC) = 0 resulting in a multiple-in-sentence-pronoun.
5. Sentence-distance(NE, SC) = 1, but recency(NE) = 2, resulting in a short-distance-pronoun.
6. Similarly, LRRH6 is converted into a short-distance-pronoun.
7. As with element #4, this is a case resulting in a multiple-in-sentence-pronoun.
9. As with element #5, this is a case resulting in a short-distance-pronoun.
10. The first mention of Hugh10 in the narrative resets its discourse history entry.
11. As with element #2, the discourse plan reports that this is an appositive.
13. Finally, Hugh13 is repeated in the same sentence.

Figure 2: A Brief Trace of the Pronominalization Algorithm for Anaphoric Pronouns from S TORY B OOK

narrative (discourse) plan into a multi-page fic- one of two techniques:

tional narrative in the fairy tale domain. Using a
pipelined generation architecture, S TORY B OOK per-  Checking algorithms by hand: One verification
forms pronominalization before sentence planning, method is to manually examine a text, identify-
and includes a revision component that is sensitive ing candidates for pronominalization and simu-
to pronominalization choices during clause aggre- lating the rules of a particular theory. However,
gation. A previous large-scale evaluation of S TORY- this method is prone to human error.
B OOK (Callaway and Lester, 2001a) which included
both a full version and a version with the pronomi-  Checking algorithms semiautomatically: Other
nalization component ablated showed that including researchers opt instead to annotate a corpus
such a component significantly increases the quality for pronominalization and their antecedents as
of the resulting prose. well as the pronoun forms that should occur,
and then simulate a pronominalization algo-
However, there are significant practical obstacles rithm on the marked-up text (Henschel et al.,
to comparing the performance of different pronomi- 2000). Similarly, this approach can suffer from
nalization algorithms using corpus matching criteria interannotator agreement errors (Poesio et al.,
instead of “quality” as evaluated by human judges. 1999b).
Because systems that can handle a large quantity of
text are very recent and because it can require years To verify our pronominalization algorithm more
to create and organize the necessary knowledge to rigorously, we instead used the S TORY B OOK deep
produce even one multi-paragraph text, much re- generation system to recreate pre-existing multi-
search on anaphora generation has instead relied on page texts with automatically selected pronouns.
 McCoy & Strube Henschel et al. S TORY B OOK S TORY B OOK
NYT News NYT News NYT News LRRH Narrative
Animate Anaphora 370/437 (84.7%) N/A 415/449 (92.4%) 170/174 (97.7%)
All Anaphora N/A 469/527 (89.0%) 441/475 (92.8%) 177/181 (97.8%)
Cataphora N/A N/A 1/2 (50.0%) 1/2 (50.0%)
Dialogue N/A N/A 46/46 (100.0%) 65/65 (100.0%)
Deixis N/A N/A 9/9 (100.0%) None present
Reflex./Recip. N/A N/A 5/6 (83.3%) 2/2 (100.0%)
Partitive N/A N/A 1/2 (50.0%) 1/1 (100.0%)

Table 1: Pronouns Correct by Algorithm/Text vs. Pronoun Type

Without a full-scale implementation, it is impossible texts. Previous approaches, based largely on theo-
to determine whether an algorithm performs imper- retical approaches such as Centering Theory, deal
fectly due to human error, a lack of available corpus exclusively with anaphoric pronouns and have com-
data for making decisions, or if it is a fault with the plex processing and definitional requirements.
algorithm itself. Given the full rhetorical structure available to an
Using the algorithm described in Figure 1, we implemented generation system, we devised a sim-
modified S TORY B OOK to substitute the types of pler method of determining appropriate pronom-
pronouns described in Section 3. We then created inalizations which was more accurate than exist-
the discourse plan and lexicon necessary to generate ing methods simulated by hand or performed semi-
the same three articles from the New York Times as automatically. This shows that approaches designed
(McCoy and Strube, 1999). The results for both the for use with anaphora resolution, which must build
newspaper texts and the Little Red Riding Hood nar- up discourse knowledge from scratch, may not be
rative described in (Callaway and Lester, in press) the most desirable method for use in NLG, where
are shown in Table 1. discourse knowledge already exists. The positive re-
With the same three texts from the New York sults from our simple counting algorithm, after only
Times, S TORY B OOK performed better than the pre- minor changes in parameters from a narrative do-
vious reported results of 85-90% described in (Mc- main to that of newspaper text, indicates that future
Coy and Strube, 1999; Henschel et al., 2000) on both high-quality prose generation systems are very near.
animate and all anaphora using a corpus matching
technique. Furthermore, this was obtained solely by 8 Acknowledgements
adjusting the recency parameter to 4 (it was 3 in our We would like to thank Michael Young and Renate
narrative domain), and without considering other en- Henschel for their helpful comments; Kathy McCoy
hancements such as gender/number constraints or very quickly provided the original 3 NYT articles
domain-specific alterations.1 upon request; the anonymous reviewers whose com-
ments greatly improved this paper. Support for this
7 Conclusions work was provided by ITC-irst and the IntelliMedia
Pronominalization is an important element in the au- Initiative of North Carolina State University.
tomatic creation of multi-paragraph and multi-page
1
It is important to note, however, that our counts of pronouns References
and antecedents do not match theirs. This may stem from a vari-
ety of factors, such as including single instances of nominal de- Douglas E. Appelt. 1985. Planning English referring
scriptions, whether dialogue pronouns were considered, and if expressions. Artificial Intelligence, 26:1–33.
borderline quantifiers and words like “everyone” were counted.
The generation community to-date has not settled on standard, Frederick Baldwin. 1995. CogNIAC: A Discourse Pro-
marked corpora for comparison purposes as has the rest of the cessing Engine. Ph.D. thesis, The University of Penn-
computational linguistics community. sylvania, Philadelphia, PA.
Charles B. Callaway and James C. Lester. 2001a. Eval- Kathleen F. McCoy and Michael Strube. 1999. Taking
uating the effects of natural language generation on time to structure discourse: Pronoun generation be-
reader satisfaction. In Proceedings of the Twenty- yond accessibility. In Proceedings of the Twenty-First
Third Annual Conference of the Cognitive Science So- Conference of the Cognitive Science Society, pages
ciety, pages 164–169, Edinburgh, UK. 378–383, Vancouver, CA, August.

Charles B. Callaway and James C. Lester. 2001b. Nar- Elena Not. 1996. A computational model for generating
rative prose generation. In Proceedings of the Seven- referring expressions in a multilingual application do-
teenth International Joint Conference on Artificial In- main. In COLING–1996: Proceedings of the 16th In-
telligence, pages 1241–1248, Seattle, WA. ternational Conference on Computational Linguistics,
Copenhagen, Denmark, August.
Charles B. Callaway and James C. Lester. 2003. Narra-
tive prose generation. Artificial Intelligence. In press. M. Poesio, H. Cheng, R. Henschel, J. Hitzeman, R. Kib-
ble, and R. Stevenson. 1999a. Specifying the parame-
Wim Claassen. 1992. Generating referring expressions ters of centering theory: A corpus-based evaluation us-
in a multimodal environment. In R. Dale, E. Hovy, ing text from application-oriented domains. In Book-
D. Rosner, and O. Stock, editors, Aspects of Auto- title, page Pages, Address, Month.
mated Natural Language Generation, pages 247–62.
Springer-Verlag, Berlin. M. Poesio, R. Henschel, J. Hitzeman, R. Kibble, S. Mon-
tague, and K. van Deemter. 1999b. Towards an anno-
Robert Dale. 1992. Generating Referring Expressions. tation scheme for noun phrase generation. In Bookti-
MIT Press. tle, page Pages, Address, Month.
Barbara Di Eugenio. 1998. Centering in Italian. In R. Quirk, S. Greenbaum, G. Leech, and J. Svartvik. 1985.
Marilyn A. Walker, Aravind K. Joshi, and Ellen F. A Comprehensive Grammar of the English Language.
Prince, editors, Centering in Discourse. Oxford Uni- Longman Publishers.
versity Press, Cambridge, MA.
Ehud Reiter and Robert Dale. 1997. Building ap-
Barbara J. Grosz and Candace L. Sidner. 1986. Atten- plied natural-language generation systems. Journal of
tion, intentions, and the structure of discourse. Com- Natural-Language Engineering, 3:57–87.
putational Linguistics, 12(3):175–204.
Candace L. Sidner. 1983. Focusing in the com-
Barbara J. Grosz, Aravind K. Joshi, and Scott Weinstein. prehension of definite anaphora. In M. Brady and
1995. Centering: A framework for modelling the lo- R. Berwick, editors, Computational Models of Dis-
cal coherence of discourse. Computational Linguis- course, pages 267–330. MIT Press, Cambridge, MA.
tics, 21(2).
Marilyn A. Walker. 1998. Centering, anaphora resolu-
Barbara J. Grosz. 1977. The representation and use of tion, and discourse structure. In Marilyn A. Walker,
focus in a system for understanding dialogs. In Pro- Aravind K. Joshi, and Ellen F. Prince, editors, Center-
ceedings of the Fifth International Joint Conference on ing in Discourse. Oxford University Press, Cambridge,
Artificial Intelligence, pages 67–76, Cambridge, MA. MA.

Peter Heeman and Graeme Hirst. 1986. Collaborating Bonnie Webber. 1979. A Formal Approach to Discourse
on referring expressions. Computational Linguistics, Anaphora. Garland, NY.
12(3):351–382.
Maria Wolters and Donna K. Byron. 2000. Prosody and
Renate Henschel, Hua Cheng, and Massimo Poesio. the resolution of pronominal anaphora. In COLING–
2000. Pronominalization revisited. In COLING– 2000: Proceedings of the 18th International Con-
2000: Proceedings of the 18th International Con- ference on Computational Linguistics, Saarbruecken,
ference on Computational Linguistics, Saarbruecken, Germany.
Germany.
C. Yeh and C. Mellish. 1997. An empirical study on
Jerry R. Hobbs. 1976. Pronoun resolution. Technical the generation of anaphora in Chinese. Computational
Report 76-1, Department of Computer Science, City Linguistics, 23(1):169–190.
College, CUNY, New York, NY.

Roger Kibble and Richard Power. 2000. An inte-

grated framework for text planning and pronominali-
sation. In Proceedings of the First International Con-
ference on Natural Language Generation, pages 194–
200, Mitzpe Ramon, Israel.