International Conference on Advanced Language Processing and Web Information Technology

Using Semantic Roles for Coreference Resolution

2026 IEEE Green Technologies Conference
Computer Science and Engineering
Green University of Bangladesh .,Dhaka
Email:kamrulhssn6@[Link]

Abstract played by an NP (e.g. Bean and Riloff, 2004), and

In this paper, we systematically explore the use of statistics-based semantic compatibility information
semantic roles in coreference resolution. Here, the (e.g. Yang,2005). Other researchers try to apply some
semantic roles are automatically determined using a syntactic structured information into pronoun
state-of-the-art SRL system, ASSERT, and integrated resolution. Shane Bergsma and Dekang Lin (2006)
into the resolution by combining them with detailed presented an approach to pronoun resolution based on
pronoun categories. In particular, the anaphoricity of syntactic paths. Through a simple bootstrapping
the pronoun “it” in “it” sentence pattern is determined procedure, the approach achieved the likelihood of
using its semantic role and applied in a filter process. coreference between a pronoun and a candidate noun
Experimental results on the ACE 2003 corpus show based on the path in the parse tree between the two
that our method can significantly outperform the entities. Yang et al. (2006) proposed a kernel-based
baseline system by 3.1% in F-measure. method that can automatically mine the syntactic
information from the parse trees. They utilized the
1. Introduction parse trees directly as a structured feature and apply
kernel functions to this feature. Zhou et al. (2008)
proposed a context-sensitive convolution tree kernel
Coreference resolution is the process of linking for pronoun resolution. For a pair of an anaphor and an
together multiple expressions of a given entity. The antecedent candidate, they implemented a dynamic-
key to solve this problem is to determine the expansion scheme to automatically determine a proper
antecedent for each referring expression in a document. tree span by taking predicate- and antecedent
Now coreference resolution has been proven to be a competitor-related information into consideration.
major obstacle in building robust systems for Then they applied a context-sensitive convolution tree
information extraction, text summarization, question kernel for pronoun resolution.
answering and so on. So far, there are still some problems. Whether a
In the last decade, knowledge-learn approaches given noun phrase is anaphoric or non-anaphoric is
have significantly influenced the research of very important for subsequent resolution. Consider the
coreference resolution. Current coreference resolution following instances extracted from ACE 2003:
systems are mostly relying on lexical and syntactic (1) The rebel Sudan1 People 's Liberation
attribute of the noun phrases, such as string matching, Army2 said Thursday that it3 will extend by
the distance between the coreferent expressions, name another three months the current three-
alias, and so on (Soon et al.,2001;Ng and Cardie,2002). month cease-fire when it end Wednesday to
However, there is much more knowledge in the small, allow relief food to reach civilians in war
annotated corpus. And deeper linguistic knowledge provinces like Bahr el-Ghazal in the south .
needs to be made available to resolver in order to reach (2) It is destiny, he kept repeating to no avail .
the next level of performance. As a result, researchers In example 1, the pronoun “it3” is anaphoric, but in
have re-adopted the once-popular knowledge-rich example 2, the pronoun “It” is non-anaphoric. How to
approach. Some researchers focus on investigating a distinguish between these two phenomena is still a
variety of semantic knowledge source for common hard work. In this paper, we present an approach using
noun resolution, such as the semantic relations between semantic role knowledge to resolve this problem. For
two NPs (e.g. Ji et al.,2005), the contextual role the example 2, the result of semantic role labeling is:

* QianPeide: Corresponding Author

978-0-7695-3273-8/08 $25.00 © 2008 IEEE 150

DOI 10.1109/ALPIT.2008.60

Authorized licensed use limited to: Soochow University. Downloaded on December 11, 2009 at 00:29 from IEEE Xplore. Restrictions apply.
 It's destiny, [ARG0 he] kept [TARGET repeating] does not consider other relations outside the above
to no avail. three. Another problem is that there is no approach,
We can see that the pronoun “It” is a non-anaphoric. which can be used to extract semantic relationships
Another contribution of our work lies in combining automatically.
semantic role information with detailed pronoun type As for semantic class of an NP, some approaches
knowledge in coreference resolution. This combination simply assign to a common noun the first (i.e., most
can use semantic role knowledge effectively. frequent) WordNet sense as its semantic class (e.g.,
The rest of this paper is organized as follows. Soon et al.(2001), Markert and Nissim (2005)). Ng
Section 2 briefly describes some related work that (2007) acquired semantic class knowledge from a
utilizes the semantic knowledge for resolution. Section version of the Penn Treebank in which the noun
3 introduces the framework, as well as the baseline phrases were labeled with their semantic class. Ng
feature space. Section 4 presents in detail the semantic adopted a corpus-based approach to semantic class
role features. Section 5 reports and discusses the determination. He classified an NP as belonging to one
experimental results, and finally, section 6 gives the of the ACE semantic classes, so the task of semantic
conclusion. class determination can be known as a six-class
classification task.
The above discussion suggests that semantic
2. Related Work
information can significantly improve the resolution.
So far, there are still much semantic information may
Semantic information is an important factor.
not be well utilized. So the task of this paper is to find
Related work on exploring semantic information in
out whether a learning based coreference resolver can
coreference resolution can be typically classified into
be improved by using semantic role knowledge that is
three categories: computing semantic similarity
automatically acquired by some semantic role labeling
between two NPs, contextual role played by NP,
tools, and how to use semantic role knowledge
semantic class of an NP.
effectively.
So far, there are several coreference resolution
systems using WordNet, HowNet etc. to compute the
semantic similarity between two NPs. Vieira & 3. The Resolution Framework
Poesic(2000), Harabagiu et al.(2001), Poesio et
al.(2004), Markert & Nissim(2005) and Our correference resolution system adopts the
Wikipedia(2006) explore the use of WordNet for common learning-based framework similar to the one
coreference resolution. By means of a WordNet search, by Soon et al.(2001).
the results in terms of semantic similarity measures can
be used as features for a learner. Table 1: Feature set of baseline system
ANPronoun 1 if anaphor is a pronoun; else 0
As a representative for using contextual role played
ANDefiniteNP 1 if anaphor is a definite NP; else 0
by NP in coreference resolution, Yang (2005)
1 if anaphor is a demonstrative NP; else
considered three relationships: possessive-noun, ANDemonstrativeNP 0
subject-verb and verb-object. Before resolution a large CAPronoun if candidate is a pronoun; else 0
corpus was prepared. Documents in the corpus were 1 if candidate and anaphor agree in
processed by a shallow parser that could generate ANCAGenderAgreement
gender; else 0
predicate-argument tuples. During resolution, for an ANCANumberAgreemen 1 if candidate and anaphor agree in
encountered anaphor, each of its antecedent candidates t number; else 0
was substituted with the anaphor. According to the role ANCAAppositive
1 if candidate and anaphor are in an
and type of the anaphor in its context, a predicate- appositive structure; else 0
argument tuple was extracted. Each extracted tuple was 1 if candidate and anaphor match in
ANCAHeadStringMatch headword; else 0
searched in the prepared tuples set of the corpus, and
the times the tuple occurs was calculated. Because Distance between candidate and
ANCASentDistance
anaphor in sentences
corpus-based knowledge usually suffers from data
1 if candidate and anaphor agree in the
sparseness problem, Yang also obtained the predicate- ANCAWORDSENSE sense achieved from WordNet; else 0
argument statistics via a web search engine like Google 1 if candidate and anaphor are both
and Altavista. Then, the statistics from corpus and web ANCABothProperName proper name; else 0
combined with twin-candidate model to compute and 1 if candidate and anaphor are in an
apply the semantic information. Yang’s study shows ANCANameAlias alias of the other; else 0
that the semantic information can significantly improve
the resolution of neutral pronouns. However, Yang

151

Authorized licensed use limited to: Soochow University. Downloaded on December 11, 2009 at 00:29 from IEEE Xplore. Restrictions apply.
 During training, all possible markables in a training is realized in Un. Cp(Un) is the highest-ranked
document are determined by a pipeline of NLP member of Cf(Un), and is predicted to be Cb(Un+1).
components, and training examples in the form of The ranking of Cf based on Brennan, Friedman,
feature vectors which are generated for appropriate and Pollard (1987) is Subject > Object(s) > Other. That
pairs of markables. These training examples are then is to say discourse entities realized in subject position
given to SVM to build a classifier. During testing, all is ranked more highly than entities realized in object
markables are determined by the same pipeline, and position, which are both then ranked more highly than
potential pairs of noun phrase‘s markables are entities realized in subordinate clauses or as other
presented to the classifier, which decides whether the grammatical functions. So far, the ranking of Cf by
two noun phrase corefer. grammatical role has been widely adopted in many
As many other learning-based approaches, the NLP applications. However, for English, there is still a
knowledge for the reference determination is problem of the utility of grammatical role. We must
represented as a set of features associated with the consider voice. Using the active voice or using the
training or test instances. In our baseline system, we passive voice, the grammatical role of the same entity
only select features that can be obtained with low should be different. So in this paper, we apply
annotation cost and high reliability. And these features semantic role into resolution.
must be generic enough to be used across different
domains. All features are listed in Table 1 together 4.2 Semantic Role
with their prospective possible values. They are same
with the features of Soon(2001). In our research, we A semantic role is the underlying relationship that a
duplicate the Soon system as a baseline system. participant has with the main verb in a clause.
Semantic role is the actual role a participant plays in
4. Incorporating Semantic Role some real or imagined situation, apart from the
linguistic encoding of those situations.
If, in some real or imagined situation, someone
4.1 Centering Theory
named John purposely hits someone named Bill, then
John is the agent and Bill is the patient of the hitting
Centering theory is a theory of local discourse
event. Therefore, the semantic role of Bill is the same
structure that models the interaction of referential
(patient) in both of the following sentences:
continuity and salience of discourse entities in the
⚫ John hit Bill.
internal organization of a text. The main assumptions
⚫ Bill was hit by John.
of the theory as presented by Grosz, Joshi and
In both of the above sentences, John has the
Weinstein (1995) and Walker, Joshi and Prince (1998)
semantic role of agent.
can be summarized as follows:
During the last few years there has been increasing
⚫ For each utterance in a discourse there is
interest in Semantic Role Labeling (SRL). It is
precisely one entity that is the center of
currently a well defined task with a substantial body of
attention.
work and comparative evaluation. Given a sentence,
⚫ For consecutive utterances within a discourse,
the task consists of analyzing the propositions
segment should keep the same entity as the
expressed by some target verbs and some constituents
center of attention.
of the sentence. In particular, for each target verb
⚫ The entity most prominently realized in an
(predicate) all the constituents in the sentence which
utterance should be identified as the center of
fill a semantic role of the verb have to be recognized.
attention.
Typical semantic roles include Agent, Patient,
⚫ The center of attention is the entity that is most
Instrument, etc. and also adjuncts such as Locative,
likely to be pronominalized.
Temporal, Manner, and Cause, etc. Among semantic
The main claims of centering theory are formalized
roles, agent and patient are steady. Knowledge about
in terms of Cb, the backward-looking center; Cf, a list
other roles (e.g. Locative, Temporal) can be expressed
of forward-looking centers for each utterance Un; and
as Word sense or Semantic class.
Cp or preferred center, the most salient candidate for
In this paper, we conjecture that candidate is mostly
subsequent utterances. Cf(Un) is a partial ordering on
an antecedent of anaphor if it is the center of one
the entities mentioned (or “realized”) in Un, ranked by
utterance and the ranking is Agent > Patient > Others.
grammatical role; for example, SUBJ > DIR-OBJ
Our approach introduces some semantic role
>INDIR-OBJ > COMP(S) > ADJUNCT(S). Cb(Un) is
features to describe the semantic role of anaphor and
defined as the highest-ranked member of Cf(Un-1) that
candidate. Considering the stability of semantic role,
our research only focused on the most important

152

Authorized licensed use limited to: Soochow University. Downloaded on December 11, 2009 at 00:29 from IEEE Xplore. Restrictions apply.
semantic roles (Arg0 and Arg1), other roles and some same Hidden Markov Model (HMM) based engine
information about predicate-argument tuples were not with error-driven learning capability (Zhou and Su,
concerned. Semantic role features are listed in table 2. 2000 & 2002) which achieves precision and recall rate
Semantic role were labeled based on the results of a of 96.49 and 96.99 for noun phrases of Penn WSJ
chunk tagger. We used ASSERT as the SRL system,, TreeBank . The recognition of NEs as well as their
developed by Sameer Pradhan and others at University semantic categories was trained for the MUC NE task
of Colorado. and obtained high F-scores of 96.9% (MUC-6) and
94.3%(MUC-7)(Zhou and Su, 2002). Throughout the
Table 2: Semantic Role Features experiments, default learning parameters were applied
ANArg0 1 if the semantic role of anaphor is arg0 else 0 to the SVM algorithm. We report performance in terms
ANArg1 1 if the semantic role of anaphor is arg1 else 0 of recall, precision, and F-measure using the
CAArg0 1 if the semantic role of candidate is arg0 else 0 commonly-used model theoretic MUC scoring
CAArg0 1 if the semantic role of candidate is arg1 else 0 program (Vilain et al., 1995).

4.3 Detailed Pronoun Type Information 5.2 Results and Discussion

In order to use semantic knowledge effectively, we In our study, we want to know whether a learning-
applied some detailed pronoun type information into based coreference resolver can be improved using
resolution. Based on centering theory and SRL, we semantic role knowledge and how to use the semantic
know that (1) the center of attention is the entity that is role knowledge effectively. Table 4 summarizes the
most likely to be pronominalized; (2) Arg0 and Arg1 results using different feature groups.
have higher ranking than other semantic roles. So we
think semantic role knowledge should be useful for Table 4: Results of different feature groups
pronoun resolution. Detailed pronoun type feature can Features R P F
describe the pronoun more exactly, and which should Duplicated Soon Baseline 55.0 53.5 54.2
be helpful for applying semantic role knowledge. Table +Semantic Role Features 52.7 51.9 52.3
3 presents all the detailed pronoun type features. +Detailed Pronoun Type Features 54.0 51.9 52.9
+Semantic Role Features
Table 3: Detailed Pronoun Type Features & +Detailed Pronoun Type Features 53.4 60.8 56.9
1 if anaphor is a first personal +Filter the non-anaphoric “it” 53.3 58.0 55.6
ANFirstPersonPronoun
pronoun; else 0 +Semantic Role Features
1 if anaphor is a second personal & +Detailed Pronoun Type Features 53.4 61.8 57.3
ANSecondPersonPronoun pronoun; else 0 & +Filter the non-anaphoric “it”
1 if anaphor is a third personal
ANThirdPersonPronoun
pronoun; else 0 We can see that only introducing semantic role
1 if candidate is a first personal features or detailed pronoun type features, recall and
CAFirstPersonPronoun pronoun; else 0 precision of the system are all lowered, and F-measure
1 if candidate is a second are also lowered. But the combination of semantic role
CASecondPersonPronoun personal pronoun; else 0
and detailed pronoun type features can improve the
1 if candidate is a third personal
CAThirdPersonPronoun system. The F-measure is 2.7% higher than baseline
pronoun; else 0
system. Furthermore, we analyze “it” sentence pattern
based on semantic role and apply a non-anaphor filter
5. Experimentation and Discussion of “it” into resolution. The last two lines of Table 4
show that the filter of “it” is very useful. The
5.1 Experimental Setup combination of semantic role, detailed pronoun type
and filter of non-anaphoric “it” can yield a statistically
The experiments were done on ACE 2003 NWIRE significant improvement of 3.1% in F-measure.
corpus. For the training set, there are totally 4773 Arg0 The second line of Table 4 shows that only applying
and 8351 Arg1. While for the testing set, the number is semantic role features will lower the performance.
1379 and 2416. Semantic role emphasizes the center of one utterance,
An input raw text was preprocessed automatically and the center is the entity that is most likely to be
by a pipeline of NLP components, including pronominalized. But the baseline system does not
tokenization and sentence segmentation, named entity distinguish the types of pronoun. Only applying
recognition, part-of-speech tagging and noun phrase semantic role knowledge can introduce some noise.
chunking. Among them, named entity recognition, Figure 1 gives out an example.
part-of-speech tagging and text chunking apply the

153

Authorized licensed use limited to: Soochow University. Downloaded on December 11, 2009 at 00:29 from IEEE Xplore. Restrictions apply.
 In this example, the phrases “the authorities”, Table 5 lists the resolution results for five types of
“their” and “they” lies on same coreference chain. In noun phrases. As shown, using semantic role and
baseline system, the pronoun “their” and “they” can be detailed pronoun type features, the system can
resolved correctly. But using semantic role knowledge significantly boost the performance of the baseline for
separately, the system can not identify the relationship pronoun resolution.
between “they” and “their”. Based on semantic role,
for the verb “search”, the role of “they” is Arg0. And ⚫ Result of Baseline System：
for the verb “send”, the role of “they” is also Arg0. At Acting in <COREF ID="32" REF="13" >his</COREF>
same time, the pronoun “their” has no semantic role. own defense during <COREF ID="33" REF="13"
So the relationship between “their” and “they” is lost. >his</COREF> pre-sentencing hearing Monday , <COREF
In the same way, applying detailed pronoun type ID="34" REF="7" >Akayesu</COREF> said remorsefully ,
I regret what happened from the bottom of my heart .
information separately is useless, by contraries, the
⚫ Result using Semantic Role and Detailed Pronoun Type
knowledge will introduce some noise. Features
Acting in <COREF ID="32" REF="13" >his</COREF>
<s> <COREF ID = " 3053 " REF = " 3055 " > Mascola own defense during <COREF ID="33" REF="13"
</COREF> said <COREF ID = " 2441 " REF = " 2445 " > >his</COREF> pre-sentencing hearing Monday , <COREF
the authorities </COREF> may stop <COREF ID = " 2446 " ID="34" REF="7" >Akayesu</COREF> said remorsefully ,
REF = " 2441 " > their </COREF> practice of detaining <COREF ID="35" REF="13" >I</COREF> regret what
<COREF ID = " 3365 " REF = " -1 " > people </COREF> happened from the bottom of <COREF ID="36" REF="13"
for hours while <COREF ID = " 2447 " REF = " 2446 " > >my</COREF> heart .
they </COREF> search a threatened building and instead Figure 2: a segment from ACE2003 NWIRE
send <COREF ID = " 3366 " REF = " 3365 " > people
</COREF> home if there is no clear evidence of danger .
In English, there are many different use of “it”. E.g.
</s>
“This is a new car. I bought it yesterday.“ The
Figure 1: a segment from ACE2003 NWIRE
pronoun “it“ is an anaphor, the antecedent is the noun
Incorporating semantic role knowledge with phrase “a new car“. But in some “it” sentence pattern,
detailed pronoun type information, the center of one “it” is non-anaphoric. E.g. “It is destiny, he kept
utterance is emphasized, and at the same time, the type repeating to no avail.” The neutral pronoun “it” should
of pronoun will be matched. The errors of mismatch be a non-anaphor. Because its semantic role is neither
person pronoun will be eliminated. So the performance arg0 nor arg1. The last two lines show that the filter of
is improved. Consider the example described in figure non-anaphoric “it” can improve the performance.
2.
We are also concerned about how the semantic role
knowledge works for different types of noun phrases.

Table 5: resolution results for different types of noun phrases

Using Semantic Role and Detailed
Baseline System Pronoun Type Features
Type of NP Changes
R P F R P F
Proper Name 80.8 82.5 81.7 80.8 82.5 81.7 0
Pronoun 59.2 59.3 59.2 66.5 59.1 62.6 +3.4
UnDefiniteNP 33.8 40.3 36.8 41.5 31.7 36.0 -0.8
DefiniteNP 34.7 45.6 39.4 35.8 44.5 39.6 +0.2
DemonstrativeNP 90.9 10.0 18.0 90.9 10.0 18.0 0

information is used to filter out the non-anaphoric “it”.

6 Conclusion This further improve the F-measure by 0.4%.
For the future work, we will explore more kinds of
This paper systematically explores the use of semantic information in coreference resolution.
semantic role information in coreference resolution.
This is done in two aspects. First, semantic role
information is determined using an existing SRL 7 Acknowledgement
system and integrated with detailed pronoun type
information. Evaluation on the ACE 2003 corpus This research is supported by Project 60673041
shows that such integration improves the performance under the National Natural Science Foundation of
by 2.7% in F-measure. Second, the semantic role China, project 2006AA01Z147 under the “863”

154

Authorized licensed use limited to: Soochow University. Downloaded on December 11, 2009 at 00:29 from IEEE Xplore. Restrictions apply.
 National High-Tech Research and Development of Yang X.F.， Zhou G.D.， Su J. and Tan C.L. (2003).
China, and project 20060285008 under the National Coreference Resolution Using Competition Learning
Approach ， ACL ’ 2003: 176- 183 ， Sapporo ，
Research Foundation for the Doctoral Program of
Higher Education of China. Japan， 7- 12 July 2003.
[11] Yang X.F., Su J., Zhou G.D. and Tan C.L. (2004).
Improving pronoun resolution by incorporating
8 References coreferential information of candidates. ACL ’
2004:127- 134, Barcelona, Spain, 21- 26 July 2004 .
[1] Brennan, Susan E, Marilyn W. Friedman, and Carl J. [12] Zhou G.D. and Su J. (2004). A high- performance
Pollard (1987). “A centering approach to pronouns”. coreference resolution system using a multi- agent
Proceedings of the 25th Annual Meeting of the strategy， COLING’2004:522- 528. 23 - 27 Aug，
Association for Computational Linguistics, Cambridge, 2004， Geneva， Switzerland.
Mass., 290-2.
[2] Grosz, Barbara J., and Candace L. Sidner (1986). [13] Yang X.F.， Su J. and Tan C.L. (2005), Improving
“Attentions, intentions and the structure of discourse”. Pronoun Resolution Using Statistics - Based Semantic
Computational Linguistics, 12: 175-204. Compatibility Information, ACL ’ 2005:165 -172 ，
[3] Brennan, Susan E (1995) “Centering attention in Univ of Michgan-Ann Arbor， USA， 25 -30 June
discourse”. Language and Cognitive Process, 10/2: 2005
137-67. [14] David Westbrook, Ralph Grishman; Using semantic
[4] Aone C and Bennett W.W. (1995). Evaluating Relations to Refine Coreference Decisions
automated and manual acquisition of anaphora Computational Linguistics, 2005
resolution strategies. ACL’1995:122-129. [15] Zhou G.D., Su J. Zhang J. and Zhang M. (2005).
[5] Grosz, A. JoShi and S. Weinstein. 1995. Centering: a Exploring various knowledge in relation extraction.
framework for modeling the local coherence of ACL’2005. 427-434. 25-30 June, Ann Arbor, Michgan,
discourse. Computational Linguistics, 21(2):203-225. USA.
[6] Mitkov R. (1998). Robust pronoun resolution with [16] Bergsma S. and Lin D.K.(2006). Bootstrapping path-
limited knowledge. COLING-ACL’1998:869-875. based pronoun resolution. COLING-ACL’2006: 33-40.
Montreal, Canada. [17] Yang X.F., Su J. and Tan C.L. (2006). Kernel-based
[7] Collins M. (1999) Head-driven statistical models for pronoun resolution with structured syntactic
natural language parsing. Ph.D. Thesis. University of knowledge. COLING-ACL’2006: 41-48.
Pennsylvania. [18] Yang X.F. and Su J. (2007). Coreference Resolution
[8] Rodger Kibble, Goldsmiths College. (2001). A Using Semantic Relatedness Information from
Reformulation of Rule 2 of Centering. Computational Automatically Discovered Patterns ACL’2007 528-535
Linguistics, 2001,27(4): 579-587 Prague, Czech Republic, June 2007.
[9] Soon W.M., Ng H.T. and Lim. (2001). A machine [19] Ng V. (2007). Semantic Class Induction and
learning approach to coreference resolution of noun Coreference Resolution ACL’2007 536-543 Prague,
phrase. Computational Linguistics, 2001,27(4):521544. Czech Republic, June 2007.
[10] Ng V. and Cardie C. (2002). Improving machine [20] Ng, V. (2007). Shallow semantics for coreference
learning approaches to coreference resolution. resolution. In IJCAI 2007, pages 1689–1694.
Proceedings of the 40th Annual Meeting of the [21] Zhou G.D., Kong F. and Zhu Q.M. 2008a. Context-
Association for Computational Linguistics, 2002. sensitive convolution tree kernel for pronoun
resolution. IJCNLP’2008

155

155iew publicAatiuonthstoatrsized licensed use limited to: Soochow University. Downloaded on December 11, 2009 at 00:29 from IEEE Xplore. Restrictions