USING RELATIONS TO INTERPRET ANAPHORA

Parma Nand and Wai Yeap

School of Computer and Mathematical Sciences, Auckland University of Technology, Auckland, New Zealand

Keywords:

Anaphora resolution, Noun phrase anaphora, Discourse structure, Noun compounds, Noun phrases.

Abstract:

In this paper we present a novel framework for resolving bridging anaphora. The new framework is based on

the core set of relations that have been used to describe an entirely different linguistic process, the process of

generating a compound noun from two different nouns. We argue that the linguistic processes of compound

noun generation and the use of NP anaphora are alike hence have to use the same relational framework. We

validated this hypothesis by using human annotators to interpret indirect anaphora from naturally occurring

discourses. The annotators were asked to classify the relations between anaphor-antecedent pairs into relation

types that have been previously used to describe the relations between a modiﬁer and the head noun of a

compound noun. We obtained very encouraging results with a Fleiss’s κ value of 0.66 for inter-annotation

agreement.

1 INTRODUCTION

The term anaphora resolution refers to identiﬁcation

of the real world entity that an anaphor co-refers to.

This is true in the case of pronouns where the pro-

noun has a one-to-one relation with the antecedent.

It is also true in the case of some NPs where the

anaphoric NP directly co-refers to the antecedent (eg.

James Smith/Mr Smith), however in a case such as

window/house, table/tablecloth or even car/vehicle,

there is an indirect relation, different to a one-to-one

co-reference type relation between the anaphor and

the antecedent. NP anaphora resolution studies (e.g.

(Poesio et al., 1997; Fraurud, 1990; Vieira and Poe-

sio, 2000)) treat these indirect relations as a single

category and refer to them as associative or bridging

anaphora, however they are still interpreted as a form

co-reference. In this study we propose a relational

framework that distinguishes between the different

types of semantic relations that can exist between two

nouns, one of which is used anaphorically in a dis-

course. Hence anaphora interpretation involves iden-

tifying the antecedent as well as the type of relation

to the antecedent noun. Since we are now identifying

the type of relation, it is now possible for an anaphor

to have multiple antecedents, related by the same or

a different relation. This is a signiﬁcant departure

from the conventional notion of anaphora resolution

where an anaphor is resolved to a single previously

mentioned entity, and in the case in which it is also an

anaphor, it is assumed to be already resolved form-

ing a chain of relations. For some NP anaphora this

is inadequate. As an illustration, consider the excerpt

below:

The robber jumped out of the window

The house

belonged to Mr Smith.

The window

is thought to have been un-

locked.

If we allow a single resolution relation for an

anaphoric NP, then window

would have to be re-

solved to either house

or window

. In either case, a

part ofthe informationwould not be captured. A com-

mon strategy in most studies (eg. (Poesio et al., 1997;

Fraurud, 1990)) is to resolve to the most recent an-

tecedent. In the case of the above excerpt, this would

mean that we resolve window

to house

which can

be assumed to be already resolved to window

. There

are two inadequacies in this strategy; ﬁrstly the se-

mantic difference between the relation of window

window

and window

to house

is approximated by

a single co-reference relation, and secondly as a con-

sequence, the direct relation between window

to win-

dow

is not captured. In the proposed framework, we

will identify both house

and window

as antecedents

and interpret each of them with a different relation. It

can be argued that this can be overwhelming since we

can form a relation between even a pair of very remote

entities. However the constraint in our case is that

we are only interested in relations that give rise to

anaphoric use of NPs. The interpretation framework

involves specifying a relation between an anaphor

Nand P. and Yeap W..

USING RELATIONS TO INTERPRET ANAPHORA.

DOI: 10.5220/0003740700920099

In Proceedings of the 4th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2012), pages 92-99

ISBN: 978-989-8425-95-9

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

and the antecedent hence a consequence of this is that

an NP can form relations with more than one an-

tecedent. This allows us to interpret and represent the

multiple relations between a simple anaphoric noun

such as window to house and another occurrence of

window. However, more importantly, it allows us to a

richer interpretation of compound nouns by interpret-

ing the modiﬁer/s in addition to the head noun. As

a simple example, the compound noun battle fatigue,

appearing after the clause “The battle caused fatigue”

has a co-referential relation to the noun fatigue, but in

addition it also has some interpretation relation (iden-

tiﬁed later as CAUSE ) to the noun battle.

Hence, there are two novel aspects to this frame-

work for interpreting anaphora; ﬁrstly it identiﬁes

a speciﬁc relation between the anaphor and its an-

tecedent. Secondly, it also interprets modiﬁers be-

yond using them to merely identify the antecedent for

the head noun, ie. interprets them in the same way as

the head noun. A consequential effect of this is that

an NP can have more then one antecedent. Thus this

framework enables us to determine the relational de-

pendence of an anaphoric NP to all other NPs in the

discourse.

2 RELATED WORKS

NP anaphora resolution has received considerably

less attention from computational linguists compared

to pronominal anaphora even though the proportion

of NP anaphora in natural discourses is either com-

parable or more then the proportion of pronominal

anaphora. The reason for this seems to be because

the problem of pronoun resolution much better de-

ﬁned compared to NP anaphora. This complexity

of the problem also explains why whatever published

work is available on NP anaphora resolution, is pre-

dominantly focussed on NPs that are deﬁnite descrip-

tions (eg. (Poesio et al., 1997; Fraurud, 1990; Bean

and Riloff, 1999; Bunescu, 2003)) with the accom-

panying task of identifying whether a deﬁnite NP

is anaphoric or not. NP resolution in these stud-

ies involves identifying a single previously mentioned

noun that the anaphoric NP refers to. There are two

categories of anaphora; direct and associative. The

direct category includes cases in which an NP di-

rectly co-refers to another entity such as the case of

he/John. The associative category includes cases such

as window/house. Some of the studies such as (Poesio

et al., 1997) have gone a step further to specify the

actual associative relation in terms of synonymy

, hy-

same meaning relation

ponymy

and meronymy

. The motivation for these

relations seems to have risen from organization of the

lexicon, WordNet (Fellbaum, 1998) which is used to

bridge the meanings between the anaphor and the an-

tecedent.

This paper argues that the noun compound gener-

ation process and the anaphoric use of NPs are similar

linguistic processes, hence they have to be based on a

common relational framework.

A compound NP of the form noun + noun (N +

N) consists of two nouns which have some underly-

ing semantic relation ((Levi, 1973; Downing, 1977;

Li, 1971)). According to these studies use of com-

pound NPs is highly productive rather then lexical.

In this productive process, compound NPs are formed

by a producer on the ﬂy as a discourse is being pro-

duced rather then recalled and used from a lexicon.

This formation of NPs is not totally unconstrained, in

other words, a compound NP cannot be formed with

any two random nouns. For example, war man can

not be formed on the basis of the relation “man who

hates war” or similarly house tree can not be formed

from “tree between two houses” (Zimmer, 1971). In

both the examples there does exist a relation between

the nouns, however it is of the type that can be used

to form a compound NP. With the exception of (Zim-

mer, 1971), the other mentioned studies on compound

NPs have assumed that the set of generic relations are

ﬁnite and characterizable, although the set is not nec-

essarily common among all the studies. Studies such

as (Levi, 1973) and (Downing, 1977) have attempted

to identify these relations, and even though the exact

set of relations proposed by the different studies are

slightly different, a core set is very similar. An addi-

tional aspect highlighted in (Downing, 1977) is that

compound nouns can also be formed from “tempo-

rary or fortuitous” relations, hence it presents a case

for an existence of unbounded number of relations al-

though the vast majority of noun compounds ﬁt into

a relatively small set of categories (Tratz and Hovy,

2010).

The relational frameworks used in computational

linguistics vary along similar lines as those proposed

by linguists. Some works in the computational lin-

guistics (eg. (Butnariu and Veale, 2008; Nakov,

2008)) assume the existence an unbounded number

of relations while others (eg. (Lauer, 1995; Kim and

Nakov, 2011)) use categories similar to Levi’s ﬁnite

set. Yet others (eg. (Nastase and Szpakowicz, 2003;

Kim and Baldwin, 2005)) are somewhat similar to

(Warren, 1978). Most of the research to date has been

domain independent done on generic corpus such as

same subset/superset relation

part/whole relation

USING RELATIONS TO INTERPRET ANAPHORA

Penn Tree Bank, British National Corpus or the web.

The later works on noun compounds have fol-

lowed on from either (Levi, 1978) or (Warren, 1978)

with some of them coming up with a slightly differ-

ent variation while others have deﬁned a ﬁner grained

set of relations dictated by the data sets used for the

study. For example, (Tratz and Hovy, 2010) reports

a set of 43 relations grouped into 10 upper level cat-

egories. Most of the relations from different studies

can be mapped to an equivalent relation in other stud-

ies.

For this study we chose the set of relations pro-

posed by in (Levi, 1978) for two reasons. Firstly, our

analysis of corpus for anaphor-antecedent relations

seemed to map better to Levi’s set of nine relations

for noun compounds and secondly more of these rela-

tions can be computationally determined from exist-

ing lexicons such as WordNet and the Web. In terms

of natural language processing, a linguistic theory is

only useful if it can be reasonable implemented in a

computational system.

(Levi, 1978) proposes that compound NPs are de-

rived from underlying clause or complement struc-

tures by the two processes of predicate deletion and

normalization. Her work is based on similar frame-

work as (Lees, 1966) except Less’ transformational

process is based on verb classiﬁcations. Levi pro-

poses that, in the case of normalization, ‘the underly-

ing predicate survives overtly in the head noun, with

the modiﬁer deriving from either the subject or the

object of the underlying S’, giving rise to subjective

(eg. industrial production) or objective (eg. heart

massage) normalization. For the case of predicate

deletion, Levi maintains that the number or deletable

predicates is limited to only nine primitive relations.

They are CAUSE, HAVE, MAKE, USE, BE, IN,

FROM, ABOUT and FOR. In the next section we ar-

gue that these relations also describe anaphoric use of

NPs.

3 ANAPHORA RESOLUTION

FRAMEWORK

In the Introduction we stated that anaphora interpreta-

tion and noun compound generation are two indicants

of the same underlying relational framework between

entities. Hence, a framework describing compound

noun generation has to apply to anaphora usage as

well. In the proposed framework we extend the re-

lations proposed for compound noun generation by

predicate deletion from (Levi, 1978) for interpretation

of NP as well as pronominal anaphora.

An indirect reference such as window referring to

house and diesel referring to truck is based on the

predicates “house has windows” and “a truck uses

diesel”. In the case of noun compound generation,

the predicate is deleted and the two entities are jux-

taposed to form the noun compounds house window

and diesel truck. For interpretation of the compound

noun the consumer is expected to reconstruct the rela-

tion between the modiﬁer and the head noun ((Down-

ing, 1977; Levi, 1978)). We propose that the com-

pound noun generation process is very similar to NP

anaphora, except in the latter case the modiﬁer is not

necessarily bound to the head noun as part of a noun

compound. That is, it may exist in another clause,

however the same relation is still expected to be re-

constructed for a full interpretation of the anaphor.

Hence, for the example for the predicate “house has

window”, we could have the full NP, house window

produced by predicate deletion. However in addition,

the same predicate coulc also be expressed anaphori-

cally as in the following example:

John bought a house in Glen Eden.

The windows are wooden.

In the example above the related entities from the

predicate “house has windows” are separated into two

different sentences, however the consumer is still ex-

pected to reconstruct the semantic relation as in the

case of noun compound generation. Hence, with the

proposed anaphora interpretation framework, the only

difference between noun compound generation and

anaphoric use of NPs is that in the former the two

nouns are used together as a compound noun while

in the latter the nouns are used separately as anaphor-

antecedent pairs, however they are both governed by

the same semantic framework.

Semantic relations between certain entities exit

by default and can be assumed as part of the lexical

knowledge by a producer. For example, the HAVE

relation between car and tyre is part of lexicon so

the noun compound car tyre and the noun tyre used

anaphorically to refer to car is readily understood.

However, in addition, a semantic relation can also be

formed between two between two entities which are

not usually related by the relation. In this case the re-

lation is explicitly expressed by a predicate as context

before the predicate is deleted to form a noun com-

pound or used anaphorically. For example, after spec-

ifying the relation “the box has tyres”, the noun tyres

can be used to indirectly refer to box in the same way

as the reference of tyres to car. However, the former

can only be used in the context of the discourse in

which the relation was expressed. This corresponds

to Downing’s (Downing, 1977) fortuitous relations.

We distinguish between these two type of relations

as persistent or contextual. Persistent relations are

ICAART 2012 - International Conference on Agents and Artificial Intelligence

those that form part of the lexical knowledge which

are valid within the context of an individual discourse

as well as all other discourses. On the other hand con-

textual relations are transient and may be valid only

for the duration of a single discourse, for example, “a

cup on a table” or “John has a knife”. The contex-

tual relations can be expressed as either a verb or a

preposition relating two entities in the discourse be-

ing processed.

As argued earlier, the semantic relations used by

NP anaphora are same as those used for noun com-

pound generation, hence for this study we adopted the

set proposed in (Levi, 1978) consisting of the rela-

tions CAUSE, HAVE, MAKE, USE, BE, IN, FOR,

FROM and ABOUT. In order to deﬁne a complete

framework for anaphora interpretation, we needed to

do two modiﬁcations to the nine relations from (Levi,

1978). Both of these modiﬁcations were done in or-

der to be able to better interpret and represent plural

anaphoric NPs. This was done by introducing a new

relation named ACTION, and by splitting the existing

BE relation into BE-INST and BE-OCCR. These are

explained next.

When two or more entities in a discourse are par-

ticipating in the same or similar event, they can be

referred to as a unit by a collective NP in the context

of the discourse. Two entities in the same or simi-

lar action can be expressed by the conjunction and

or described by two different clauses. For instance

in the sentence “The coastguard and Lion Foundation

Rescue helicopter were called out.”, the entities coast-

guard and Lion Foundation Rescue helicopter are re-

lated to each other by the virtue of participating in the

same action. Similarly, the clauses “the truck rolled

down the hill” and “the ball rolled down the hill”

would enforce the same relation between truck and

ball since they are both engaged in the same action

(roll). This relation between truck and ball is only

valid for the context of the discourse, hence the re-

lation is contextual. We describe this contextual re-

lation as the ACTION relation which relates entities

participating in events which are identiﬁed to be same

or similar. The ACTION is used to describe an NP

such as runners used to refer to fox and Peter from

the context clause “The fox and Peter were running”.

The second modiﬁcation involved deﬁning a ﬁner

grained BE relation in order to interpret existence of

plurals in a different form. We split Levi’s BE re-

lation into BE-OCCR and BE-INST to distinguish

between direct co-reference or identity relation and

an instance relation. In a BE-OCCR relation an

NP directly forms a one-to-one co-reference to an-

other NP, eg. John/he and John/the driver. The

BE-INST relation represents cases where an anaphor

refers to a plural antecedent, in a partial capacity,

for example, both trucks/northbound truck. In this

case the NP northbound truck is an instance of both

trucks which is distinct from a co-reference relation.

It can be argued that all subset/ superset relations

such as John/driver(John is an instance of driver) and

car/vehicle (car is and instance of vehicle) is an in-

stance relation. However we consider these as BE-

OCCR relation since they function to identify the en-

tity. Hence in the framework, the BE-INST relation

only relates a plural NP and an NP representing a sub-

set of the plural NP.

With this discussion we can now deﬁne and

exemplify the eleven relation types was used for the

purpose of anaphora interpretation. They are:

CAUSE - Includes all causal relations. For example,

battle/fatigue, earthquake/tsunami

HAVE - Includes notions of possession. This

includes diverse examples such as snake/poison,

house/window and cake/apple.

MAKE - Includes examples such as concrete house,

tar/road and lead/pencil.

USE - Some examples are drill/electricity and

steam/ship.

BE-INST - Includes plural cases such as both

trucks/southbound truck, John/teachers.

BE-OCCR - Describes the same instance participat-

ing in multiple events. For example John Smith/Mr

Smith/he and John Smith/the driver.

IN - This relation captures grouping of things that

share physical or temporal properties. For example

lamp/table and Auckland/New Zealand.

FOR - This includes purpose of one entity for an-

other. For example pen/writing and soccer ball/play.

FROM - This includes cases where one entity is

derived from another. For example olive/oil and

wheat/ﬂour.

ABOUT - Describes cases where one entity is a topic

of the other. For example travel/story and loan/terms.

ACTION - This is only a contextual relation meant

to capture entities engaged in same or similar action

either with the same object/s or a null object.

The next section describes the annotation experi-

ment done in order to validate the that anaphora usage

is based on the above relation types.

4 ANNOTATION EXPERIMENT

4.1 Annotators

For the purpose of human validation of all relations in

USING RELATIONS TO INTERPRET ANAPHORA

the framework we used second and third year stu-

dents enrolled in computer related degrees. The anno-

tation experiments were done overa period of 4 weeks

at the beginning of their usual classes. Four different

streams were used each consisting of approximately

30 students. The students in each stream were given

a basic training on the requirements of the annotation

and they were given a single annotation task at the

beginning of their class over a 4 week period. The

whole annotation experiment was broken down into

session based tasks involving 25 anaphoric NPs per

task. This was done to ensure that each task was com-

pleted in about 10 minutes with minimal impact on

the students class time. In addition, the annotators

were not identiﬁed in any of the tasks. We only en-

sured that an annotator did not annotate the same task

twice.

4.2 Annotation Data

Our base input data used for content analysis for all

aspects of NP usage consisted of 120 articles (of

mixed genre) from The New Zealand Herald, The Do-

minion Post and The Press which are three major on-

line newspapers from three different cities in New

Zealand. The choice of the articles were not com-

pletely random. This corpora was developed to serve

as the input data for the anaphora resolution system

which is the parent project of this study. Hence, the

corpora was developed from the articles which were

not too short (had more then 20 sentences), exhibited

use of a variety of anaphoric uses (including pronom-

inal anaphora) and had been written by different writ-

ers.

An inherent challenge in most NLP tasks is what

is referred to as data sparseness. The term is used

to describe a characteristic when a single chosen cor-

pus cannot be used for consistent empirical validation

of all aspects of a theory. This is because the preva-

lence of the different characteristics of an NLP theory

can be unevenly distributed in a ﬁxed corpus. Hence,

we searched an extended corpus in order to make the

lower threshold of 15 relations from each category.

For this we used The Corpus of Contemporary Amer-

ican English (Davies, 2010). This freely available

corpora consisting of some 410 million words from

a variety of genre and has an online web interface

which can be used to do fairly complex searches for

words and phrases hence forms an excellent resource

for manual content analysis for NLP tasks.

We excluded validating the BE-OCCR relation

since this is a non-ambiguous co-reference relation.

For the annotation experiment we used 3 streams

of approximately 30 students giving us a total of 90

different annotators. Each annotator took 4 differ-

ent tasks, one per week for a period of 4 weeks.

Each task consisted of 25 antecedent-anaphor pairs

and was annotated by 2 streams, ie. approx. 60

annotators. We randomly discarded some annota-

tion task sheets in order to have a consistent number

of annotations for each pair resulting in 25 annota-

tors for each task. Each relation type from (CAUSE,

HAVE, MAKE, USE, BE-INST, IN, FOR, FROM,

ABOUT, ACTION) as classiﬁed by the author was

represented by 15 anaphor-antecedent pairs. The

pairs from each of the 10 relation types were ran-

domly selected to make up 6 task sheets, each con-

sisting of 25 pairs. The total number of classiﬁcations

for all relations amounted to 3750 with 375 classiﬁca-

tions for each relation type consisting of 15 different

anaphor-antecedent pairs.

Each of the streams were given a basic training

on on semantic interpretation of the relation types us-

ing the examples in section section 3. These exam-

ples were also given as a separate sheet with each an-

notation task. Each task sheet consisted of anaphor-

antecedent pairs and a tick box for each of the rela-

tions. The annotators were asked to choose the re-

lation which best describes the anaphor-antecedent

pair. Two additional options, OTHER and NONE

were also given. The OTHER was to be used if the

annotator thought that a relation does exist but is not

present in the given list and option NONE to be used

if the annotator thought that the pair were not related

at all.

Table 1 shows the confusion matrix of the rela-

tion types as identiﬁed by the annotators against the

author’s classiﬁcation. Table 2 shows the correspond-

ing confusion indices between the relation types. The

confusion indices indicate the likelihood of a relation

type to be interpreted as another type.

5 ANNOTATION RESULTS

The ﬁrst observation of the annotation results from

table 1 is that only 2 annotations out of a total of

3750 were classiﬁed as NONE indicating that the

annotators by and large thought that the pairs given

had some relation. In addition a total of 43 (approx.

1.1%) annotations were classiﬁed as OTHER, indi-

cating the relations were described by a relation not in

the list of 10 that were given. The main categories that

were interpreted as having some other relation were

CAUSE and ACTION, however these were still a very

small percentage with indices of 0.04 and 0.05 respec-

tively. The bolded entries in table 2 give the percent-

age agreement of the relation types agreeing with that

ICAART 2012 - International Conference on Agents and Artificial Intelligence

Table 1: Confusion Matrix for the non-normalized NPs. The columns give the annotations by annotators against the author’s

annotations on the rows. Each relation category had a total of 375 annotations done by 50 different annotators. The bolded

entries indicate number of annotations agreeing with author’s annotations.

CAUSE HAVE MAKE USE BE-INST IN FOR FROM ABOUT ACTION OTHER NONE

CAUSE 208 45 87 4 14 15 2

HAVE 196 113 7 13 46

MAKE 45 206 120 4

USE 45 26 242 59 3

BE-INST 347 19 9

IN 64 37 241 33

FOR 18 5 132 216 4

FROM 9 17 35 87 227

ABOUT 48 11 56 7 253

ACTION 5 351 19

Table 2: Confusion Index Matrix between the relation types corresponding to table 1. The bolded entries indicate the index

of annotations which were the same as that of the author’s.

CAUSE HAVE MAKE USE BE-INST IN FOR FROM ABOUT ACTION OTHER NONE

CAUSE 0.55 0.12 0.23 0.01 0.04 0.04 0.01

HAVE 0.52 0.30 0.02 0.03 0.12

MAKE 0.12 0.55 0.32 0.01

USE 0.12 0.07 0.65 0.16 0.01

BE-INST 0.93 0.05 0.02

IN 0.17 0.10 0.64 0.09

FOR 0.05 0.01 0.35 0.58 0.01

FROM 0.02 0.05 0.09 0.23 0.61

ABOUT 0.13 0.03 0.15 0.02 0.67

ACTION 0.00 0.01 0.94 0.05

of the author. The relation types BE-INST and AC-

TION have the highest conformance indicating they

are the least ambiguous. The other types vary from

a low ﬁgure of 0.52 for HAVE to 0.67 for ABOUT

with an overall agreement value of 0.66. The rela-

tion types that were easily confused and hence can

be interpreted as semantically close were HAVE and

MAKE USE. Conﬂating these 3 categories gives us

an agreement index of 0.89. Another crucial observa-

tion is for the FROM relation. Although not by large

amounts, this relation type seems to be confused with

all other categories. This prompted us to closely ex-

amine the task sheets to see if there were consistent

misclassiﬁcations by the author, however no such pat-

terns were found. Some of the classiﬁcations seemed

to use the FROM type as a “fall back” category.

In order to compare the inter-annotator agree-

ment with other similar studies we also computed the

Fleiss’ κ measure. The κ index for the overall anno-

tation tasks was computed to be 0.64 and the value

with HAVE, MAKE and USE conﬂated was 0.86.

The overall κ value 0.64 compares well the inter-

annotator ﬁgures from other annotation experiments

dealing with identiﬁcation of relations. For compari-

son some of the results are summarized in table 3.

Table 3: Inter-annotator agreement comparison between

studies dealing with relations between composite nouns of

noun compounds.

Study Agreement Index No. of Relations

(Tratz and Hovy, 2010) 0.57 - 0.67 κ 43

(Girju, 2007) 0.61 κ 22

(

O S´eaghdha, 2007) 0.68 κ 6

(Kim and Baldwin, 2005) 52.31 % 20

(Girju et al., 2005) 0.58 κ 21

6 DISCUSSION

The annotation experiment results strongly indicates

that the the two natural language usage phenomena

of compound noun generation and anaphoric use of

nouns are based on the same underlying semantic

structure. Since they are two different processes they

have been studied as two separate nlp research areas

as well. The proposed framework for anaphora reso-

lution allows us to marry the two nlp areas so that we

can better share computational advances in the two

research areas. Recently there has been an increased

momentum (Kim and Nakov, 2011; Tratz and Hovy,

2010; Nakov, 2008; Butnariu and Veale, 2008; Kim

and Baldwin, 2006; Nastase et al., 2006; Barker and

Szpakowicz, 1998; Hendrickx et al., 2010) towards

USING RELATIONS TO INTERPRET ANAPHORA

automatic derivation of relations between composite

nouns in noun compounds, most of them based on re-

lations from (Levi, 1978). This will result in an in-

creasing amount of ontology describing the semantic

relations between compound nouns which will also

become useful for anaphora interpretation if there is

an existence of a common framework. The anaphora

interpretation framework proposed in this study is a

step towards this.

Another signiﬁcant advantage of a common

framework is that it will be easier to integrate the full

meaning of a compound noun and the meaning as-

sociated with it being used anaphorically. Currently,

anaphora is described using a different set of relations

(eg. synonymy, hypernomy, meronomy etc.) and

compound nounswith adifferentset. Hence, when in-

terpreting a compound noun which is also anaphoric,

it becomes difﬁcult to merge the two meanings. Com-

bining the processes within the same framework gives

us a much stronger interpretativepower enabling us to

interpret a modiﬁer as well as the head noun. As an

illustration consider the excerpt below:

John’s car had an accident yesterday. Its

thought faulty car tyres played a major role

in the accident.

The compound noun faulty car tyres expresses the

relation HAVE between the modiﬁer car and the head

noun tyres, deﬁned by the compound noun genera-

tion framework. In terms of straight forward anaphora

resolution, the noun compound faulty car tyres is not

anaphoric since the head noun tyres does not co-refer

to anything in the previous sentence. However to be

able to fully interpret the meaning of the second sen-

tence, it is crucial that we know that the noun car in

the ﬁrst sentence also has a HAVE relation to tyres

in the second sentence. This relation forms the basis

of the coherence between “car and accident” in the

ﬁrst sentence and “tyres and accident” in the second

sentence. The proposed framework enables us to use

relations from the same set to describe the relations

between car and tyres in the compound noun faulty

car tyres and the anaphoric relations between faulty

car tyres and car. The resultant output from process-

ing a whole discourse using the proposed framework

would be a network of entity-to-entity relations con-

sisting of all freely existing nouns as well as nouns

participating as modiﬁers. This network can either be

used on its own or used as a building block towards

higher level discourse structures such as a coherence

structure.

7 CONCLUDING REMARKS

In this paper we presented a relational framework for

interpreting anaphoric NPs which goes beyond the

conventional co-reference relations. We argued that

anaphora usage and compound noun generation are

based on a common relational framework. To support

this we used an existing NP production framework

and validated it for anaphora usage using real world

discourses. We also argued that using this framework,

a more accurate level of discourse interpretation can

be achieved, both directly, as well as using it as a

building block for a higher level discourse structure

such as the coherence structure. We are in the process

of enthusiastically implementing the framework and

will be reporting the results in near future. It is antic-

ipated that successful computation of this framework

will help in numerousNLP tasks such as document vi-

sualization, summarization, archieving/retrieval and

search engine applications.

REFERENCES

Barker, K. and Szpakowicz, S. (1998). Semi-automatic

recognition of noun modiﬁer relationships.

Bean, D. and Riloff, E. (1999). Corpus-based identiﬁca-

tion of non-anaphoric noun phrases. In Proceedings of

the 37th annual meeting of the Association for Com-

putational Linguistics on Computational Linguistics,

pages 373–380. Association for Computational Lin-

guistics Morristown, NJ, USA.

Bunescu, R. (2003). Associative anaphora resolution:

A web-based approach. In In Proceedings of the

EACL2003 Workshop on the Computational Treat-

ment of Anaphora, pages 47–52.

Butnariu, C. and Veale, T. (2008). A concept-centered ap-

proach to noun-compound interpretation. In Proceed-

ings of the 22nd International Conference on Compu-

tational Linguistics - Volume 1, COLING ’08, pages

81–88, Stroudsburg, PA, USA. Association for Com-

putational Linguistics.

Davies, M. (2010). Corpus of contemporary american en-

glish. http://www.americancorpus.org/.

Downing, P. (1977). On the creation and use of english

compound nouns. Language, 53(4):810–842.

Fellbaum, C. (1998). WordNet: An Electronic Lexical

Database. Bradford Books.

Fraurud, K. (1990). Deﬁniteness and the processing of noun

phrases in natural discourse. Journal of Semantics,

7(4):395–433.

Girju, R. (2007). Improving the interpretation of noun

phrases with crosslinguistic information. In in Pro-

ceedings of the 45th Annual Meeting of the Associa-

tion of Computational Linguistics, pages 568–575.

Girju, R., Moldovan, D., Tatu, M., and Antohe, D. (2005).

On the semantics of noun compounds. Computer

Speech and Language, 19(4):479–496.

ICAART 2012 - International Conference on Agents and Artificial Intelligence

Hendrickx, I., Kim, S. N., Kozareva, Z., Nakov, P.,

S´eaghdha, D. O., Pad´o, S., Pennacchiotti, M., Ro-

mano, L., and Szpakowicz, S. (2010). Semeval-2010

task 8: Multi-way classiﬁcation of semantic relations

between pairs of nominals. In Proceedings of the 5th

International Workshop on Semantic Evaluation, Se-

mEval ’10, pages 33–38, Stroudsburg, PA, USA. As-

sociation for Computational Linguistics.

Kim, S. N. and Baldwin, T. (2005). Automatic interpre-

tation of noun compounds using wordnet similarity.

In In Proceedings of the 2nd International Joint Con-

ference on Natural Language Processing, Jeju Island,

South Korea, 1113, pages 945–956.

Kim, S. N. and Baldwin, T. (2006). Interpreting seman-

tic relations in noun compounds via verb semantics.

In Proceedings of the COLING/ACL on Main confer-

ence poster sessions, COLING-ACL ’06, pages 491–

498, Stroudsburg, PA, USA. Association for Compu-

tational Linguistics.

Kim, S. N. and Nakov, P. (2011). Large-scale noun com-

pound interpretation using bootstrapping and the web

as a corpus. In The Proceeding of Conference on

Empirical Methods in Natural Language Processing,

EMNLP., Edinburgh, UK.

Lauer, M. (1995). Corpus statistics meet the noun com-

pound: some empirical results. In Proceedings of

the 33rd annual meeting on Association for Compu-

tational Linguistics, ACL ’95, pages 47–54, Strouds-

burg, PA, USA. Association for Computational Lin-

guistics.

Lees, Robert, B. (1966). On a transformational analysis of

compounds: A reply to hans marchand. Indogerman-

ische Forschungen, pages 1–13.

Levi, J. (1973). Where do all those other adjectives come

from. In Chicago Linguistic Society, volume 9, pages

332–354.

Levi, J. N. (1978). The syntax and semantics of complex

nominals / Judith N. Levi. Academic Press, New York

Li, C. N. (1971). Semantics and the Structure of Com-

pounds in Chinese. PhD thesis, University of Carli-

fornia dissertation.

Nakov, P. (2008). Noun compound interpretation using

paraphrasing verbs: Feasibility study.

Nastase, V. and Szpakowicz, S. (2003). Exploring noun-

modiﬁer semantic relations. In Proceedings of the 5th

International Workshop on Computational Semantics.

Nastase, V. S.-S., Sokolova, J., and M. Szpakowicz, S.

(2006). Learning noun-modiﬁer semantic relations

with corpus-based and wordnet-based features. Pro-

ceedings of the National Conference on Aritﬁcial In-

telligence., 21(Part 1):781–787.

O S´eaghdha, D. (2007). Annotating and learning com-

pound noun semantics. In Proceedings of the 45th

Annual Meeting of the ACL: Student Research Work-

shop, ACL ’07, pages 73–78, Stroudsburg, PA, USA.

Association for Computational Linguistics.

Poesio, M., Vieira, R., and Teufel, S. (1997). Resolving

bridging references in unrestricted text. In Proceed-

ings of a Workshop on Operational Factors in Prac-

tical, Robust Anaphora Resolution for Unrestricted

Texts, pages 1–6, Morristown, NJ, USA. Association

for Computational Linguistics.

Tratz, S. and Hovy, E. (2010). A taxonomy, dataset, and

classiﬁer for automatic noun compound interpretation.

In Proceedings of the 48th Annual Meeting of the

Association for Computational Linguistics, ACL ’10,

pages 678–687, Stroudsburg, PA, USA. Association

for Computational Linguistics.

Vieira, R. and Poesio, M. (2000). An empirically based

system for processing deﬁnite descriptions. Computa-

tional Linguistics, 26(4):539–593.

Warren, B. (1978). Semantic patterns of noun-noun com-

pounds. Gothenburg studies in English. Acta Univer-

sitatis thoburgensis.

Zimmer, K. E. (1971). Some general observations about

nominal compounds. Working Papers on Language

Universals, pages C1–21.

USING RELATIONS TO INTERPRET ANAPHORA