Proceedings of the Workshop on Frontiers in Corpus Annotation II: Pie in the Sky, pages 5–12,
Ann Arbor, June 2005. c©2005 Association for Computational Linguistics
  
Merging PropBank, NomBank, TimeBank, Penn Discourse Treebank and Coreference 
James Pustejovsky, Adam Meyers, Martha Palmer, Massimo Poesio 
 
Abstract 
Many recent annotation efforts for English 
have focused on pieces of the larger problem 
of semantic annotation, rather than initially 
producing a single unified representation. 
This paper discusses the issues involved in 
merging four of these efforts into a unified 
linguistic structure: PropBank, NomBank, the 
Discourse Treebank and Coreference 
Annotation undertaken at the University of 
Essex. We discuss resolving overlapping and 
conflicting annotation as well as how the 
various annotation schemes can reinforce 
each other to produce a representation that is 
greater than the sum of its parts. 
 
1. Introduction 
 
The creation of the Penn Treebank (Marcus et al, 
1993) and the word sense-annotated SEMCOR 
(Fellbaum, 1997) have shown how even limited 
amounts of annotated data can result in major 
improvements in complex natural language 
understanding systems. These annotated corpora 
have led to high-level improvements for parsing 
and word sense disambiguation (WSD), on the 
same scale as previously occurred for Part of 
Speech tagging by the annotation of the Brown 
corpus and, more recently, the British National 
Corpus (BNC) (Burnard, 2000). However, the 
creation of semantically annotated corpora has 
lagged dramatically behind the creation of other 
linguistic resources: in part due to the perceived 
cost, in part due to an assumed lack of theoretical 
agreement on basic semantic judgments, in part, 
finally, due to the understandable unwillingness 
of  research groups to get involved in such an 
undertaking. As a result, the need for such 
resources has become urgent.   
 
Many recent annotation efforts for English have 
focused on pieces of the larger problem of 
semantic annotation, rather than producing a 
single unified representation like Head-driven 
Phrase Structure Grammar (Pollard and Sag 
1994) or the Prague Dependency Tecto-
gramatical Representation (Hajicova & Kucer-
ova, 2002). PropBank (Palmer et al, 2005) 
annotates predicate argument structure anchored 
by verbs. NomBank (Meyers, et. al., 2004a) 
annotates predicate argument structure anchored 
by nouns.  TimeBank (Pustejovsky et al, 2003) 
annotates the temporal features of propositions 
and the temporal relations between propositions. 
The Penn Discourse Treebank (Miltsakaki et al 
2004a/b) treats discourse connectives as 
predicates and the sentences being joined as 
arguments. Researchers at Essex were 
responsible for the coreference markup scheme 
developed in MATE (Poesio et al, 1999; Poesio, 
2004a) and have annotated corpora using this 
scheme including a subset of the Penn Treebank 
(Poesio and Vieira, 1998), and the GNOME 
corpus (Poesio, 2004a).  This paper discusses the 
issues involved in creating a Unified Linguistic 
Annotation (ULA) by merging annotation of 
examples using the schemata from these efforts. 
Crucially, all individual annotations can be kept 
separate in order to make it easy to produce 
alternative annotations of a specific type of 
semantic information without need to modify the 
annotation at the other levels. Embarking on 
separate annotation efforts has the advantage of 
allowing researchers to focus on the difficult 
issues in each area of semantic annotation and 
the disadvantage of inducing a certain amount of 
tunnel vision or task-centricity – annotators 
working on a narrow task tend to see all 
phenomena in light of the task they are working 
on, ignoring other factors. However, merging 
these annotation efforts allows these biases to be 
dealt with. The result, we believe, could be a 
more detailed semantic account than possible if 
the ULA had been the initial annotation effort 
rather than the result of merging. 
 
There is a growing community consensus that 
general annotation, relying on linguistic cues, 
and in particular lexical cues, will produce an 
enduring resource that is useful, replicable and 
portable.  We provide the beginnings of one such 
level derived from several distinct annotation 
efforts. This level could provide the foundation 
for a major advance in our ability to 
automatically extract salient relationships from 
text. This will in turn facilitate breakthroughs in 
message understanding, machine translation, fact 
retrieval, and information retrieval. 
 
2. The Component Annotation Schemata 
 
We describe below existing independent 
annotation efforts, each one of which is focused 
on a specific aspect of the semantic 
representation task: semantic role labeling, 
5
  
coreference, discourse relations, temporal 
relations, etc.  They have reached a level of 
maturity that warrants a concerted attempt to 
merge them into a single, unified representation, 
ULA.  There are several technical and theoretical 
issues that will need to be resolved in order to 
bring these different layers together seamlessly.  
Most of these approaches have annotated the 
same type of data, Wall Street Journal text, so it 
is also important to demonstrate that the 
annotation can be extended to other genres such 
as spoken language.  The demonstration of 
success for the extensions would be the training 
of accurate statistical semantic taggers. 
 
PropBank: The Penn Proposition Bank focuses 
on the argument structure of verbs, and provides 
a corpus annotated with semantic roles, 
including participants traditionally viewed as 
arguments and adjuncts.  An important goal is to 
provide consistent semantic role labels across 
different syntactic realizations of the same verb, 
as in the window in [ARG0 John] broke [ARG1 
the window] and [ARG1 The window] broke. 
Arg0 and Arg1 are used rather than the more 
traditional Agent and Patient to keep the 
annotation as theory-neutral as possible, and to 
facilitate mapping to richer representations.  The 
1M word Penn Treebank II Wall Street Journal 
corpus has been successfully annotated with 
semantic argument structures for verbs and is 
now available via the Penn Linguistic Data 
Consortium as PropBank I (Palmer, et. al., 2005).   
Coarse-grained sense tags, based on groupings of 
WordNet senses, are being added, as well as 
links from the argument labels in the Frames 
Files to FrameNet frame elements.  There are 
close parallels to other semantic role labeling 
projects, such as FrameNet (Baker, et. al., 1998; 
Fillmore & Atkins, 1998; Fillmore & Baker, 
2001), Salsa (Ellsworth, et.al, 2004), Prague 
Tectogrammatics (Hajicova & Kucerova, 2002) 
and IAMTC, (Helmreich, et. al., 2004) 
 
NomBank: The NYU NomBank project can be 
considered part of the larger PropBank effort and 
is designed to provide argument structure for 
instances of about 5000 common nouns in the 
Penn Treebank II corpus (Meyers, et. al., 2004a).  
PropBank argument types and related verb 
Frames Files are used to provide a commonality 
of annotation.  This enables the development of 
systems that can recognize regularizations of 
lexically and syntactically related sentence 
structures, whether they occur as verb phrases or 
noun phrases. For example, given an IE system 
tuned to a hiring scenario (MUC-6, 1995), 
NomBank and PropBank annotation facilitate  
generalization over patterns. PropBank and 
NomBank would both support a single IE pattern 
stating that the object (ARG1) of appoint is John 
and the subject (ARG0) is IBM, allowing a 
system to detect that IBM hired John from each 
of the following strings: IBM appointed John, 
John was appointed by IBM, IBM's appointment 
of John, the appointment of John by IBM and 
John is the current IBM appointee.  
 
Coreference: Coreference involves the detection 
of subsequent mentions of invoked entities, as in 
George Bush,… he….  Researchers at Essex (UK) 
were responsible for the coreference markup 
scheme developed in MATE (Poesio et al, 1999; 
Poesio, 2004a), partially implemented in the 
annotation tool MMAX and now proposed as an 
ISO standard; and have been responsible for the 
creation of two small, but commonly used 
anaphorically annotated corpora – the Vieira / 
Poesio subset of the Penn Treebank (Poesio and 
Vieira, 1998), and the GNOME corpus (Poesio, 
2004a).   Parallel coreference annotation efforts 
funded by ACE have resulted in similar 
guidelines, exemplified by BBN’s recent 
annotation of Named Entities, common nouns 
and pronouns.   These two approaches provide a 
suitable springboard for an attempt at achieving a 
community consensus on coreference. 
 
Discourse Treebank:  The Penn Discourse 
Treebank (PDTB) (Miltsakaki et al 2004a/b) is 
based on the idea that discourse connectives are 
predicates with associated argument structure 
(for details see (Miltsakaki et al 2004a, 
Miltsakaki et al 2004b). The long-range goal is 
to develop a large scale and reliably annotated 
corpus that will encode coherence relations 
associated with discourse connectives, including 
their argument structure and anaphoric links, 
thus exposing a clearly defined level of discourse 
structure and supporting the extraction of a range 
of inferences associated with discourse 
connectives. This annotation references the Penn 
Treebank annotations as well as PropBank, and 
currently only considers Wall Street Journal text. 
 
TimeBank: The Brandeis TimeBank corpus, 
funded by ARDA, focuses on the annotation of 
all major aspects in natural language text 
associated with temporal and event information 
(Day, et al, 2003, Pustejovsky, et al, 2004). 
Specifically, this involves three areas of the 
annotation: temporal expressions, event-denoting 
6
  
expressions, and the links that express either an 
anchoring of an event to a time or an ordering of 
one event relative to another. Identifying events 
and their temporal anchorings is a critical aspect  
of reasoning, and without a robust ability to 
identify and extract events and their temporal 
anchoring from a text, the real aboutness of the 
article can be missed.  The core of TimeBank is a 
set of 200 news reports documents, consisting of 
WSJ, DUC, and ACE articles, each annotated to 
TimeML 1.2 specification. It is currently being 
extended to AQUAINT articles. The corpus is 
available from the timeml.org website. 
 
3. Unifying Linguistic Annotations 
  
Since September, 2004, researchers representing 
several different sites and annotation projects 
have begun collaborating to produce a detailed 
semantic annotation of two difficult sentences. 
These researchers aim to produce a single unified 
representation with some consensus from the 
NLP community. This effort has given rise to 
both a listserv email list and this workshop: 
http://nlp.cs.nyu.edu/meyers/pie-in-the-sky.html, 
http://nlp.cs.nyu.edu/meyers/frontiers/2005.html 
The merging operations discussed here would 
seem crucial to the furthering of this effort. 
 
3.1 The Initial Pie in the Sky Example 
 
The following two consecutive sentences have 
been annotated for Pie in the Sky.  
 
Two Sentences From ACE Corpus File 
NBC20001019.1830.0181 
 
• but Yemen's president says the FBI has told 
him the explosive material could only have 
come from the U.S., Israel or two Arab 
countries. 
• and to a former federal bomb investigator, 
that description suggests a powerful 
military-style plastic explosive c-4 that can 
be cut or molded into different shapes. 
 
Although the full Pie-in-the-Sky analysis 
includes information from many different 
annotation projects, the Dependency Structure in 
Figure 1 includes only those components that 
relate to PropBank, NomBank, Discourse 
annotation, coreference and TimeBank. Several 
parts of this representation require further 
explanation. Most of these are signified by the 
special arcs, arc labels, and nodes. Dashed lines 
represent transparent arcs, such as the transparent 
dependency between the argument (ARG1) of 
modal can and the or. Or is transparent in that it 
allows this dependency to pass through it to cut 
and mold. There are two small arc loops -- 
investigator is its own ARG0 and description is 
its own ARG1. Investigator is a relational noun 
in NomBank. There is assumed to be an 
underlying relation between the Investigator 
(ARG0), the beneficiary or employer (the ARG2) 
and the item investigated (ARG1). Similarly, 
description acts as its own ARG1 (the thing 
described). There are four special coreference arc 
labels: ARG0-CF, ARG-ANAPH, EVENT-
ANAPH and ARG1-SBJ-CF. At the target of 
these arcs are pointers referring to phrases from 
the previous sentence or previous discourse. The 
first three of these labels are on arcs with the 
noun description as their source. The ARG0-CF 
label indicates that the phrase Yemen's president 
(**1**) is the ARG0, the one who is doing the 
describing. The EVENT-ANAPH label points to 
a previous mention of the describing event, 
namely the clause: The FBI told him the 
explosive material… (**3**). However, as noted 
above, the NP headed by description represents 
the thing described in addition to the action. The 
ARG-ANAPH label points to the thing that the 
FBI told him the explosive material can only 
come from … (**2**). The ARG1-SBJ-CF label 
links the NP from the discourse what the bomb 
was made from as the subject with the NP 
headed by explosive as its predicate, much the 
same as it would in a copular construction such 
as: What the bomb was made from is the 
explosive C-4. Similarly, the arc ARG1-APP 
marks C-4 as an apposite, also predicated to the 
NP headed by explosive. Finally, the thick arcs 
labeled SLINK-MOD represent TimeML SLINK 
relations between eventuality variables, i.e.,  the 
cut and molded events are modally subordinate 
to the suggests proposition. The merged 
representation aims to be compatible with the 
projects from which it derives, each of which 
analyzes a different aspect of linguistic analysis. 
Indeed most of the dependency labels are based 
on the annotation schemes of those projects. 
 
We have also provided the individual PropBank, 
NomBank and TimeBank annotations below in 
textual form, in order to highlight potential 
points of interaction. 
 
PropBank:  and [
Arg2 
to a former federal bomb 
investigator], [
Arg0
 that description]  
[
Rel_suggest.01
 suggests]  [
Arg1 
[
Arg1
 a powerful 
military-style plastic explosive c-4] that 
7
  
 [
ArgM-MOD
 can] be [
Rel_cut.01
 cut] or  [
Rel_mold.01
 
molded] [
ArgM-RESULT
 into different shapes]]. 
 
NomBank: and to a former [
Arg2
 federal] [
Arg1
 
bomb] [
Rel
 investigator], that description 
suggests a powerful [
Arg2
 military] - [
Rel
 style] 
plastic [
Arg1
 explosive] c-4 that can be cut 
or molded into different shapes. 
 
TimeML: and to a former federal bomb 
investigator, that description [
Event = ei1
 
suggests]  a powerful military-style plastic 
explosive c-4 that  can be [
Event = ei2 modal=’can’
 cut] 
or  [
Event = ei3 modal=’can’
 molded]  into different 
shapes. <SLINK eventInstanceID = ei1 
subordinatedEventID = ei2 relType = ‘Modal’/> 
<SLINK eventInstanceID = ei1 
subordinatedEventID = ei3 relType = ‘Modal’/> 
 
  
Figure 1. Dependency Analysis of Sentence 2  
 
Note that the subordinating Events indicated by 
the TimeML SLINKS refer to the predicate 
argument structures labeled by PropBank, and 
that the ArgM-MODal also labeled by PropBank 
contains modality information also crucial to the 
SLINKS. While the grammatical modal on cut 
and mold is captured as an attribute value on the 
event tag, the governing event predicate suggest 
introduces a modal subordination to its internal 
argument, along with its relative clause. While 
this markup is possible in TimeML, it is difficult 
to standardize (or automate, algorithmically) 
since arguments are not marked up unless they 
are event denoting.  
 
3.2 A More Complex Example 
 
To better illustrate the interaction between 
annotation levels, and the importance of merging 
information resident in one level but not 
necessarily in another, consider the sentence 
below which has more complex temporal 
properties than the Pie-in-the-Sky sentences and 
its dependency analysis (Figure 2). 
 
 According to reports, sea trials for a patrol boat 
developed by Kazakhstan are being conducted 
and the formal launch is planned for the 
beginning of April this year.  
 
 
Figure 2.  Dependency Analysis of a Sentence 
with Interesting Temporal Properties 
 
The graph above incorporates these distinct 
annotations into a merged representation, much 
like the previous analysis. This sentence has 
more TimeML annotation than the previous 
sentence.  Note the loops of arcs which show that 
According to plays two roles in the sentence: (1) 
it heads a constituent that is the ARGM-ADV of 
the verbs conducted and planned; (2) it indicates 
that the information in this entire sentence is 
attributed to the reports. This loop is problematic 
in some sense because the adverbial appears to 
modify a constituent that includes itself. In 
actuality, however, one would expect that the 
ARGM-ADV role modifies the sentence minus 
the adverbial, the constituent that you would get 
if you ignore the transparent arc from ARGM-
8
  
ADV to the rest of the sentence.  Alternatively, a 
merging decision may elect to delete the ARGM-
ADV arcs, once the more specific predicate 
argument structure of the sentence adverbial 
annotation is available. 
 
The PropBank annotation for this sentence 
would label arguments for develop, conduct and 
plan, as given below. 
 
 [
ArgM-ADV
 According to reports], [
Arg1
sea trials for  
[
Arg1
 a patrol boat] [
Rel_develop.02
 developed] [
Arg0 
by Kazakhstan]] are being  
[
Rel_conduct.01
 conducted]  and [
Arg1
 the formal 
launch] is [
Rel_plan.01
 planned]  
[
ArgM-TMP
 for the beginning of April this year].  
 
NomBank would add arguments for report, trial, 
launch and beginning as follows: 
 
 According to [
Rel_report.01
 reports], [
Arg1
 [
ArgM-LOC
 
sea [
Rel_trial.01
 trials] [
Arg1
 for [
Arg1-CF_launch.01
 a 
patrol boat] developed by Kazakhstan] are being 
conducted and the [
ArgM-MNR
 formal] [
Rel_launch.01
 
launch] is planned for the [[
REL_beginning.01
 
beginning] [
ARG1
 of April this year]].  
 
TimeML, however, focuses on the anchoring of 
events to explicit temporal expressions (or 
document creation dates) through TLINKs, as 
well as subordinating relations, such as those 
introduced by modals, intensional predicates, 
and other event-selecting predicates, through 
SLINKs. For discussion, only part of the 
complete annotation is shown below.  
  
According to [
Event = ei1  
reports], sea [
Event = ei3  
trials] for a boat [
Event = ei4  
developed]  by 
Kazakhstan are being [
Event = ei5  
conducted] and 
the formal [
Event = ei6  
launch] 
 is  [
Event = ei7  
planned] for the [
Timex3= t1  
beginning 
of April] [
Timex3= t2 
this year]. 
<SLINK eventID=”ei1” subordinatedEvent=”ei5, 
ei7” relType=EVIDENTIAL/> 
<TLINK eventID=”ei4” relatedToEvent =”ei3” 
relType=BEFORE/> 
<TLINK eventID=”ei6” relatedToTime=”t1” 
relType=IS_INCLUDED /> 
<SLINK eventID=”ei7” 
subordinatedEvent=”ei6” relType=”MODAL”/> 
<TLINK eventID=”ei5” relatedToEvent=”ei3” 
relType=IDENTITY/> 
 
Predicates such as plan and nominals such as 
report are lexically encoded to introduce 
SLINKs with a specific semantic relation, in this 
case, a “MODAL” relType,. This effectively 
introduces an intensional context over the 
subordinated events. 
 
These examples illustrate the type of semantic 
representation we are trying to achieve.  It is 
clear that our various layers already capture 
many of the intended relationships, but they do 
not do so in a unified, coherent fashion.  Our 
goal is to develop both a framework and a 
process for annotation that allows the individual 
pieces to be automatically assembled into a 
coherent whole.   
 
4.0 Merging Annotations  
 
4.1 First Order Merging of Annotation 
We begin by discussing issues that arise in 
defining a single format for a merged 
representation of PropBank, NomBank and 
Coreference, the core predicate argument 
structures and  referents for the arguments.   One 
possible representation format would be to 
convert each annotation into features and values 
to be added to a larger feature structure. 
1
 The 
resulting feature structure would combine stand 
alone and offset annotation – it would include 
actual words and features from the text as well as 
special features that point to the actual text 
(character offsets) and, perhaps, syntactic trees 
(offsets along the lines of PropBank/NomBank). 
Alternative global annotation schemes include 
annotation graphs (Cieri & Bird, 2001), and 
MATE (Carletta, et. al., 1999).  There are many 
areas in which the boundaries between these 
annotations have not been clearly defined, such 
as the treatment of support constructions and 
light verbs, as discussed below.  Determining the 
most suitable format for the merged 
representation should be a top priority. 
 
4.2 Resolving Annotation Overlap 
There are many possible interactions between 
different types of annotation: aspectual verbs 
have argument labels in PropBank, but are also 
important roles for temporal relations.  Support 
                                                 
 
1
 The Feature Structure has many advantages as a target 
representation including: (1) it is easy to add lots of detailed 
features; and (2) the mathematical properties of Feature 
Structures are well understood, i.e., there are well-defined 
rule-writing languages, subsumption and unification 
relations, etc. defined for Feature Structures (Carpenter, 
1992) The downside is that a very informative Feature 
Structure is difficult for a human to read.  
 
9
  
constructions also have argument labels, and the 
question arises as to whether these should be 
associated with the support verb or the 
predicative nominal.  Given the sentence They 
gave the chefs a standing ovation, a PropBank 
component will assign role labels to arguments 
of give; a NomBank component will assign 
argument structure to ovation that labels the 
same participants. If the representations are 
equivalent, the question arises as to which of 
them (or both) should be included in the merged 
representation. The following graph  (Figure 3) 
is a combined PropBank and NomBank analysis 
of this sentence. "They" is the ARG0 of both 
"give" and "ovation"; "the chefs" is the ARG2 of 
"give", but the "ARG1" of ovation; "ovation" is 
the ARG1 of "give" and "give" is a support verb 
for "ovation". For this case, a reasonable choice 
might be to preserve the argument structure from 
both NomBank and PropBank, and to do the 
same for other predicative nominals that have 
give (or receive, obtain, request…) as a support 
verb, e.g., (give a kiss/hug/squeeze, give a 
lecture/speech, give a promotion, etc.).   For 
other support constructions, such as take a walk, 
have a headache and make a mistake, the noun is 
really the main predicate and it is questionable 
whether the verbal argument structure carries  
gave
chefsthe
They
a ovationstanding
NP
NP
S
ARG0
REL
ARG2
ARG1
NP
ARG1 REL
ARG0SUPPORT
 
Figure 3. Merged PropBank/NomBank representation 
of They gave the chefs a standing ovation. 
much information, e.g., there are no selection 
restrictions between light verbs and their subject 
(ARG0) -- these are inherited from the noun. 
Thus make a mistake selects a different type of 
subject than make a gain, e.g., people and 
organizations make mistakes, but stock prices 
make gains. For these constructions, the merged 
representation might not need to include the 
(ARG0) relation between the subject of the 
sentence and make, and future propbanking 
efforts might do well to ignore the shared 
arguments of such instances and leave them for 
NomBank. However, the merged representation 
would inherit PropBank’s annotation of some 
other light verb features including: negation, e.g., 
They did not take a walk; modality, e.g., They 
might take a walk; and sentence adverbials, e.g., 
They probably will take a walk. 
 
4.3 Resolving Annotation Conflicts 
Interactions between linguistic phenomena can 
aid in quality control, and conflicts found during 
the deliberate merging of different annotations 
provides an opportunity to correct and fine-tune 
the original layers. For example, predicate 
argument structure (PropBank and NomBank) 
annotation sometimes assumes different 
constituent structure than the Penn Treebank. We 
have noticed some tendencies that help resolve 
these conflicts, e.g., prenominal noun 
constituents as in Indianapolis 500, which forms 
a single argument in NomBank, is correctly 
predicted to be a constituent, even though the 
Penn Treebank II assumes a flatter structure.  
 
Similarly, idioms and multiword expressions 
often cause problems for both PropBank and 
NomBank. PropBank annotators tend to view 
argument structure in terms of verbs and 
NomBank annotators tend to view argument 
structure in terms of nouns. Thus many examples 
that, perhaps, should be viewed as idioms are 
viewed as special senses of either verbs or nouns. 
Having idioms detected and marked before 
propbanking and nombanking could greatly 
improve efficiency.   
 
Annotation accuracy is often evaluated in terms 
of inter-annotation consistency. Task definitions 
may need to err on the side of being more 
inclusive in order to simplify the annotators task. 
For example, the NomBank project assumes the 
following definition of a support verb (Meyers, 
et.al., 2004b):  “… a verb which takes at least 
two arguments NP
1
 and XP
2
 such that XP
2
 is an 
argument of the head of NP
1
. For example, in 
John took a walk, a support verb (took) shares 
one of its arguments (John) with the head of its 
other argument (walk).” The easiest way to 
apply this definition is without exception, so it 
will include idiomatic expressions such as keep 
tabs on, take place, pull strings. Indeed, the 
dividing line between support constructions and 
idioms is difficult to draw (Meyers 2004b).   
PropBank annotators are also quite comfortable 
with associating general meanings to the main 
verbs of idiomatic expressions and labeling their 
10
  
argument roles, as in cases like bring home the 
bacon and mince words with. Since idioms often 
have interpretations that are metaphorical 
extensions of their literal meaning, this is not 
necessarily incorrect.  It may be helpful to have 
the literal dependencies and the idiomatic 
reading both represented. The fact that both 
types of meaning are available is evidenced by 
jokes, irony, and puns.  
 
With respect to idioms and light verbs, TimeML 
can be viewed as a mediator between PropBank 
and NomBank. In TimeML, light verbs and the 
nominalizations accompanying them are marked 
with two separate EVENT tags. This guarantees 
an annotation independent of textual linearity 
and therefore ensures a parallel treatment for 
different textual configurations. In (a) the light 
verb construction "make an allusion" is 
constituted of a verb and an NP headed by an 
event-denoting noun, whereas in (b) the nominal 
precedes a VP, which in addition contains a 
second N:  
(a) Max [made an allusion] to the crime.  
(b) Several anti-war [demonstrations have taken 
place] around the globe. 
Both verbal and nominal heads are tagged 
because they both contribute relevant 
information to characterizing the nature of the 
event. The nominal element plays a role in the 
more semantically based task of event 
classification. On the other hand, the information 
in the verbal component is important at two 
different levels: it provides the grammatical 
features typically associated with verbal 
morphology, such as tense and aspect, and at the 
same time it may help in disambiguating cases 
like take/give a class, make/take a phone call. 
The two tagged events are marked as identical by 
a TLINK introduced for that purpose. The 
TimeML annotation for the example in (a) is 
provided below.  
Max [
Event = ei1  
made] an [
Event = ei2  
allusion] to 
the crime.  
<TLINK eventID="ei1"relatedToEvent="ei2" 
relType=IDENTITY> 
Some cases of support in NomBank could also 
be annotated as "bridging" anaphora. Consider 
the sentence: The pieces make up the whole. 
It is unclear whether make up is a support verb 
linking whole as the ARG1 of pieces or if pieces 
is linked to whole by bridging anaphora. 
There are also clearer cases. In Nastase, a rival 
player defeated Jimmy Connors in the third 
round, the word rival and Jimmy Connors are 
clearly linked by bridging. However, a wayward 
NomBank annotator might construct a support 
chain (player + defeated) to link rival with its 
ARG1 Jimmy Connors.  In such a case, a 
merging of annotation could reveal annotation 
errors. In contrast, a NomBank annotator would 
be correct in linking John as an argument of walk 
in John took a series of walks (the support chain 
took + series consists of a support verb and a 
transparent noun), but this may not be obvious to 
the non-NomBanker. Thus the merging of 
annotation may result in the more consistent 
specifications for all.  
 
In our view, this process of annotating all layers 
of information and then merging them in a 
supervised manner, taking note of the conflicts, 
is a necessary prerequisite to defining more 
clearly the boundaries between the different 
types of annotation and determining how they 
should fit together.  Other areas of annotation 
interaction include: (1) NomBank  and 
Coreference, e.g. deriving that John teaches 
Mary from John is Mary's teacher involves: (a) 
recognizing that teacher is an argument 
nominalization such that the teacher is the ARG0 
of teach (the one who teaches); and (b) marking 
John and teacher as being linked by predication 
(in this case, an instance of type coreference); 
and (2) Time and Modality -  when a fact used to 
be true, there are two time components: one in 
which the fact is true and one in which it is false. 
Clearly more areas of interaction will emerge as 
more annotation becomes available and as the 
merging of annotation proceeds.  
 
5. Summary 
 
We proposed a way of taking advantage of the 
current practice of separating aspects of semantic 
analysis of text into small manageable pieces. 
We propose merging these pieces, initially in a 
careful, supervised way, and hypothesize that the 
result could be a more detailed semantic analysis 
than was previously available. This paper 
discusses some of the reasons that the merging 
process should be supervised. We primarily gave 
examples involving the interaction of PropBank, 
NomBank and TimeML. However, as the 
merging process continues, we anticipate other 
conflicts that will require resolution. 
 
References 
C. F. Baker, F. Collin, C. J. Fillmore, and J. B.  
Lowe (1998), The Berkeley FrameNet 
project. In Proc. of COLING/ACL-98,  86--90   
O. Babko-Malaya, M. Palmer, X. Nianwen, S.  
Kulick, A. Joshi (2004), Propbank II, 
Delving Deeper, In Proc.  of HLT-NAACL 
Workshop: Frontiers in Corpus Annotation. 
R. Carpenter (1992), The Logic of Typed  
Feature Structures. Cambridge Univ. Press. 
J. Carletta and A. Isard (1999), The MATE  
Annotation Workbench: User Requirements. 
In Proc. of the ACL Workshop: Towards 
Standards and Tools for Discourse Tagging. 
Univ. of Maryland, 11-17 
C. Cieri and S. Bird (2001), Annotation Graphs  
and Servers and Multi-Modal Resources: 
Infrastructure for  Interdisciplinary Education, 
Research and Development Proc. of the ACL 
Workshop on Sharing Tools and Resources 
for Research  and Education, 23-30 
D. Day,  L. Ferro, R. Gaizauskas, P. Hanks, M.  
Lazo, J. Pustejovsky, R. Saurí, A. See, A. 
Setzer, and B. Sundheim (2003), The 
TIMEBANK Corpus. Corpus Linguistics. 
M. Ellsworth, K. Erk, P. Kingsbury and S. Pado  
(2004), PropBank, SALSA, and FrameNet: 
How Design Determines Product, in Proc. of  
LREC 2004 Workshop: Building Lexical 
Resources from Semantically Annotated 
Corpora.  
C. Fellbaum (1997), WordNet: An Electronic  
Lexical Database, MIT Press.. 
C. J. Fillmore and B. T. S. Atkins (1998), 
FrameNet and lexicographic relevance. In the 
Proc. of the First International Conference 
on Language Resources and Evaluation.  
C. J. Fillmore and C. F. Baker (2001), Frame  
semantics for text understanding. In Proc. of 
NAACL WordNet and Other Lexical 
Resources Workshop. 
E. Hajivcova and I. Kuvcerov'a (2002).  
Argument/Valency Structure in PropBank, 
LCS Database and Prague Dependency 
Treebank: A Comparative Pilot Study. In the 
Proc. of the Third International Conference 
on Language Resources and Evaluation 
(LREC 2002),  846--851. 
S. Helmreich, D. Farwell, B. Dorr, N. Habash, L. 
Levin, T. Mitamura, F. Reeder, K. Miller, E. 
Hovy, O. Rambow and A. Siddharthan,(2004), 
Interlingual Annotation of Multilingual Text 
Corpora, Proc. of the HLT-EACL Workshop 
on Frontiers in Corpus Annotation. 
A, Meyers, R. Reeves, C. Macleod, R, Szekely,  
V. Zielinska, B. Young, and R. Grishman  
(2004a), The NomBank Project: An Interim 
Report, Proc. of HLT-EACL Workshop: 
Frontiers in Corpus Annotation. 
A. Meyers, R. Reeves, and C. Macleod (2004b),  
NP-External Arguments: A Study of 
Argument Sharing in English. In The ACL 
2004 Workshop on Multiword Expressions: 
Integrating Processing. 
E. Miltsakaki, R. Prasad, A. Joshi and B. Webber. 
 (2004a), The Penn Discourse Treebank. In 
Proc. 4th International Conference on 
Language Resources and Evaluation (LREC 
2004). 
E. Miltsakaki, R. Prasad, A. Joshi and B. Webber  
(2004b), Annotation of Discourse 
Connectives and their Arguments, in Proc. of 
HLT-NAACL Workshop: Frontiers in Corpus 
Annotation 
M.  Marcus, B. Santorini, and M. Marcinkiewicz  
(1993), Building a large annotated corpus of 
english: The penn treebank. Computational 
Linguistics, 19:313--330. 
M. Palmer, D. Gildea, P. Kingsbury (2005), The  
Proposition Bank: A Corpus Annotated with 
Semantic Roles, Computational Linguistics 
Journal, 31:1. 
M. Poesio (2004a), The MATE/GNOME  
Scheme for Anaphoric Annotation, Revisited, 
Proc. of SIGDIAL 
M. Poesio (2004b), Discourse Annotation and  
Semantic Annotation in the GNOME Corpus, 
Proc. of ACL Workshop on Discourse 
Annotation. 
M. Poesio and M. Alexandrov-Kabadjov (2004), 
A general-purpose, off-the-shelf system for 
anaphora resol.. Proc. of LREC. 
M. Poesio, F. Bruneseaux, and L. Romary  
(1999), The MATE meta-scheme for 
coreference in dialogues in multiple language, 
Proc. of the ACL Workshop on Standards for 
Discourse Tagging.  
M. Poesio and R. Vieira (1998), A corpus-based  
investigation of definite description use. 
Computational Linguistics, 24(2). 
C. Pollard and I. A. Sag (1994), Head-driven  
phrase structure grammar. Univ. of Chicago 
Press. 
J. Pustejovsky, R. Saurí, J. Castaño, D. R. 
 Radev, R. Gaizauskas, A. Setzer, B. 
Sundheim and G. Katz (2004), Representing 
Temporal and Event Knowledge for QA 
Systems. In Mark T. Maybury (ed.), New 
Directions in Question Answering, MIT Press. 
J. Pustejovsky,  B. Ingria, R. Saurí, J. Castaño, J.  
Littman, R. Gaizauskas, A. Setzer, G. Katz, 
and I. Mani (2003), The Specification 
Language TimeML. In I. Mani, J. 
Pustejovsky, and R. Gaizauskas, editors, The 
Language of Time: A Reader. Oxford Univ. 
Press. 
