Workshop on TextGraphs, at HLT-NAACL 2006, pages 81–88,
New York City, June 2006. c©2006 Association for Computational Linguistics
Matching Syntactic-Semantic Graphs for Semantic Relation Assignment
Vivi Nastase1 and Stan Szpakowicz1,2
1 School of Information Technology and Engineering,
University of Ottawa, Ottawa, Canada
2 Institute of Computer Science,
Polish Academy of Sciences, Warsaw, Poland
{vnastase,szpak}@site.uottawa.ca
Abstract
We present a graph-matching algorithm
for semantic relation assignment. The al-
gorithm is part of an interactive text analy-
sis system. The system automatically ex-
tracts pairs of syntactic units from a text
and assigns a semantic relation to each
pair. This is an incremental learning algo-
rithm, in which previously processed pairs
and user feedback guide the process. Af-
ter each assignment, the system adds to its
database a syntactic-semantic graph cen-
tered on the main element of each pair of
units. A graph consists of the main unit
and all syntactic units with which it is syn-
tactically connected. An edge contains in-
formation both about syntax and about se-
mantic relations for use in further process-
ing. Syntactic-semantic graph matching is
used to produce a list of candidate assign-
ments for 63.75% of the pairs analysed,
and in 57% of situations the correct rela-
tions is one of the system’s suggestions;
in 19.6% of situations it suggests only the
correct relation.
1 Introduction
When analysing texts, it is essential to see how el-
ements of meaning are interconnected. This is an
old idea. The  rst chronicled endeavour to con-
nect text elements and organise connections between
them goes back to the 5th century B.C. and the work
of Panini1. He was a grammarian who analysed San-
skrit (Misra, 1966). The idea resurfaced forcefully
at several points in the more recent history of lin-
guistic research (Tesni ere, 1959; Gruber, 1965; Fill-
more, 1968). Now it has the attention of many re-
searchers in natural language processing, as shown
by recent research in semantic parsing and semantic
1The sources date his work variously between the 5th and
7th century.
role labelling (Baker et al., 1998; Kipper et al., 2000;
Carreras and Marquez, 2004; Carreras and Marquez,
2005; Atserias et al., 2001; Shi and Mihalcea, 2005).
Graph-like structures are a natural way of or-
ganising one’s impressions of a text seen from
the perspective of connections between its simpler
constituents of varying granularity, from sections
through paragraphs, sentences, clauses, phrases,
words to morphemes.
In this work we pursue a well-known and of-
ten tacitly assumed line of thinking: connections at
the syntactic level re ect connections at the seman-
tic level (in other words, syntax carries meaning).
Anecdotal support for this stance comes from the
fact that the grammatical notion of case is the basis
for semantic relations (Misra, 1966; Gruber, 1965;
Fillmore, 1968). Tesni ere (1959), who proposes a
grouping of verb arguments into actants and circum-
stances, gives a set of rules to connect speci c types
of actants  for example, agent or instrument  to
such grammatical elements as subject, direct object,
indirect object. This idea was expanded to include
nouns and their modi ers through verb nominaliza-
tions (Chomsky, 1970; Quirk et al., 1985).
We work with sentences, clauses, phrases and
words, using syntactic structures generated by a
parser. Our system incrementally processes a text,
and extracts pairs of text units: two clauses, a verb
and each of its arguments, a noun and each of its
modi ers. For each pair of units, the system builds a
syntactic graph surrounding the main element (main
clause, head verb, head noun). It then tries to  nd
among the previously processed instances another
main element with a matching syntactic graph. If
such a graph is found, then the system maps pre-
viously assigned semantic relations onto the current
syntactic graph. We have a list of 47 relations that
manifest themselves in compound clauses, inside a
simple clause or in noun phrases. The list, a syn-
thesis of a number of relation lists cited in the lit-
erature, has been designed to be general, domain-
independent (Barker et al., 1997a).
Section 2 overviews research in semantic relation
analysis. Section 3 describes the text we used in ex-
81
periments, and the semantic relation list. Section 4
looks in detail at the graph-matching heuristic. Sec-
tion 5 describes the experimental setting and shows
how often the heuristic was used when processing
the input text. We show in detail our  ndings about
syntactic levels (how often graph matching helped
assign a relation between two clauses, a verb and its
arguments, or a noun and its modi er) and about the
accuracy of the suggestion. Discussion and conclu-
sions appear in Section 6.
2 Related Work
Some methods of semantic relation analysis rely on
prede ned templates  lled with information from
processed texts (Baker et al., 1998). In other meth-
ods, lexical resources are speci cally tailored to
meet the requirements of the domain (Rosario and
Hearst, 2001) or the system (Gomez, 1998). Such
systems extract information from some types of syn-
tactic units (clauses in (Fillmore and Atkins, 1998;
Gildea and Jurafsky, 2002; Hull and Gomez, 1996);
noun phrases in (Hull and Gomez, 1996; Rosario et
al., 2002)). Lists of semantic relations are designed
to capture salient domain information.
In the Rapid Knowledge Formation Project (RKF)
a support system was developed for domain experts.
It helps them build complex knowledge bases by
combining components: events, entities and modi-
 ers (Clark and Porter, 1997). The system’s inter-
face facilitates the expert’s task of creating and ma-
nipulating structures which represent domain con-
cepts, and assigning them relations from a relation
dictionary.
In current work on semantic relation analysis, the
focus is on semantic roles  relations between verbs
and their arguments. Most approaches rely on Verb-
Net (Kipper et al., 2000) and FrameNet (Baker et
al., 1998) to provide associations between verbs and
semantic roles, that are then mapped onto the cur-
rent instance, as shown by the systems competing in
semantic role labelling competitions (Carreras and
Marquez, 2004; Carreras and Marquez, 2005) and
also (Gildea and Jurafsky, 2002; Pradhan et al.,
2005; Shi and Mihalcea, 2005).
These systems share two ideas which make them
different from the approach presented here: they all
analyse verb-argument relations, and they all use
machine learning or probabilistic approaches (Prad-
han et al., 2005) to assign a label to a new instance.
Labelling every instance relies on the same previ-
ously encoded knowledge (see (Carreras and Mar-
quez, 2004; Carreras and Marquez, 2005) for an
overview of the systems in the semantic role la-
belling competitions from 2004 and 2005). Pradhan
et al. (2005) combine the outputs of multiple parsers
to extract reliable syntactic information, which is
translated into features for a machine learning ex-
periment in assigning semantic roles.
Our system analyses incrementally pairs of units
coming from three syntactic levels  clause (CL),
intra-clause (or verb-argument, IC), noun-phrase
(NP). There are no training and testing data sets. In-
stead of using previously built resources, the system
relies on previously processed examples to  nd the
most appropriate relation for a current pair. Because
the system does not rely on previously processed
or annotated data, it is  exible. It allows the user
to customize the process for a speci c domain by
choosing the syntactic units of interest and her own
list of relations that best  t the domain.
It is also interesting to assess, using the current
system con guration, the effect of syntactic infor-
mation and incremental learning on semantic analy-
sis. This is described in section 5.
Because of these differences in the type of data
used, and in the learning approach, the results we
obtain cannot be compared to previous approaches.
In order to show that the system does learn, we show
that the number of examples for which it provides
the correct answer increases with the number of ex-
amples previously analysed.
3 Input data and semantic relations
3.1 Input data
We work with a semi-technical text on meteorolog-
ical phenomena (Larrick, 1961), meant for primary
school students. The text gradually introduces con-
cepts related to precipitation, and explains them. Its
nature makes it appropriate for the semantic analy-
sis task in an incremental approach. The system will
mimic the way in which a human reader accumu-
lates knowledge and uses what was written before to
process ideas introduced later in the text.
The text contains 513 sentences, with an average
length of 9.13 words. There are 4686 word tokens
and 969 types. The difference between the num-
ber of types (2850) and tokens (573) in the extracted
pairs (which contain only open-class words) shows
that the same concepts recur, as expected in a didac-
tic text.
The syntactic structures of the input data are
produced by a parser with good coverage and de-
tailed syntactic information, DIPETT (Delisle and
Szpakowicz, 1995). The parser, written in Prolog,
implements a classic constituency English grammar
from Quirk et al. (1985). Pairs of syntactic units
connected by grammatical relations are extracted
from the parse trees. A dependency parser would
82
produce a similar output, but DIPETT also provides
verb subcategorization information (such as, for ex-
ample, subject-verb-object or subject-verb-object-
indirect object), which we use to select the (best)
matching syntactic structures.
To  nd pairs, we use simple structural informa-
tion. If a unit is directly embedded in another unit,
we assume a subordinate relation between them; if
the two units are coordinate, we assume a coordinate
relation. These assumptions are safe if the parse is
correct. A modi er is subordinate to its head noun,
an argument to its head verb, and a clause perhaps to
the main clause in the sentence.
If we conclude that two units should interact, we
seek an appropriate semantic relation to describe this
interaction. The system uses three heuristics to  nd
one or more semantic relation candidates for the cur-
rent pair.
1. Word match  the system will propose the se-
mantic relation(s) that have previously been as-
signed to a pair containing the same lemmas.
2. Syntactic graph match  we elaborate this
heuristic in Section 4.
3. Marker  the system uses a manually built dic-
tionary of markers (prepositions, coordinators,
subordinators) associated with the semantic re-
lations they indicate. The dictionary contains
325 markers, and a total of 662 marker-relation
associations.
If neither of the three heuristics yield results, the
system will present an empty list, and expect the user
to input the appropriate relation. When at least one
relation is proposed, the user can accept a unique
relation, choose among several options, or supply
a new one. The system records which action took
place, as well as the heuristic that generated the op-
tions presented to the user. The pair is also analysed
to determine the syntactic level from which it came,
to allow for a more detailed analysis of the behaviour
of the system.
3.2 Semantic relations
The list of semantic relations with which we work
is based on extensive literature study (Barker et al.,
1997a). Three lists of relations for three syntactic
levels  inter-clause, intra-clause (case) and noun-
modi er relations  were next combined based on
syntactic and semantic phenomena. The resulting
list is the one used in the experiments we present
in this paper. The relations are grouped by general
similarity into 6 relation classes (H denotes the head
of a base NP, M denotes the modi er).
1. CAUSAL groups relations enabling or oppos-
ing an occurrence. Examples:
cause - H causes M:  u virus;
effect - H is the effect (was caused by) M:
exam anxiety;
purpose - H is for M: concert hall;
2. CONJUNCTIVE includes relations that describe
the conjunction or disjunction of occurrences
(events/act/actions/states/activities), entities or
attributes:
conjunction - both H and M occur or exist
(and nothing more can be said about that
from the point of view of causality or
temporality): running and swimming (are
good for you);
disjunction - either one or both H and M occur
or exist: painting or drawing;
3. PARTICIPANT groups relations between an oc-
currence and its participants or circumstances.
Examples:
agent - M performs H: student protest;
object - M is acted upon by H: metal separa-
tor;
bene ciary - M bene ts from H: student dis-
count;
4. SPATIAL groups relations that place an occur-
rence at an absolute or relative point in space.
Examples:
direction - H is directed towards M: outgoing
mail;
location - H is the location of M: home town;
location at - H is located at M: desert storm;
5. TEMPORAL groups relations that place an oc-
currence at an absolute or relative point in time.
Examples:
frequency - H occurs every time M occurs:
weekly game;
time at - H occurs when M occurs: morning
coffee;
time through - H existed while M existed: 2-
hour trip;
6. QUALITY groups the remaining relations be-
tween a verb or noun and its arguments. Exam-
ples:
manner - H occurs as indicated by M: stylish
writing;
83
material - H is made of M: brick house;
measure - M is a measure of H: heavy rock;
There is no consensus in the literature on a list of
semantic relations that would work in all situations.
This is, no doubt, because a general list of relations
such as the one we use would not be appropriate for
the semantic analysis of texts in a speci c domain,
such as for example medical texts. All the relations
in the list we use were necessary, and suf cient, for
the analysis of the input text.
4 Syntactic-semantic graph-matching
Our system begins operation with a minimum of
manually encoded knowledge, and accumulates in-
formation as it processes the text. This design idea
was adopted from TANKA (Barker et al., 1997b).
The only manually encoded knowledge is a dictio-
nary of markers (subordinators, coordinators, prepo-
sitions). This resource does not affect the syntactic-
semantic graph-matching heuristic.
Because the system gradually accumulates
knowledge as it goes through the input text, it uses a
form of memory-based learning to make predictions
about the semantic relation that  ts the current pair.
The type of knowledge that it accumulates consists
of previously analysed pairs, together with the
semantic relation assigned, and a syntactic-semantic
graph centered on each word in a sentence which
appears as the main element in a processed pair.
To process a pair P not encountered previously,
the system builds a graph centered on the main ele-
ment (often the head) of P. This idea was inspired
by Delisle et al. (1993), who used a list of argu-
ments surrounding the main verb together with the
verb’s subcategorization information and previously
processed examples to analyse semantic roles (case
relations). In recent approaches, syntactic informa-
tion is translated into features which, together with
information from FrameNet, WordNet or VerbNet,
will be used with ML tools to make predictions for
each example in the test set (Carreras and Marquez,
2004; Carreras and Marquez, 2005).
Our system builds a (simple) graph surrounding
a head word (which may be a verb  representing
the predicate of a sentence, or representing a clause
 or noun), and matches it with previously analysed
examples.
A graph G(w) centered on word w consists of
the following: a node for w; a set of nodes for
each of the words wi in the sentence with which w
is connected by a grammatical relation (including
situations when w is a modi er/argument); edges
that connect w with each wi, tagged with gram-
matical relation GR (such as subject, object, com-
plement) and connective information Con (preposi-
tions, coordinators, subordinators, or nil). The nodes
also contain part-of-speech information for the cor-
responding word, and other information from the
parser (such as subcategorization structure for the
verb, if it is available).
Graph matching starts with the central node, and
continues with edge matching. If G(w) is the graph
centered on word w whose pairs are currently being
processed, the system selects from the collection of
previously stored graphs, a set of graphs fG(wi)g,
which satisfy the following conditions:
 The central nodes match. The matching is
guided by a set of contraints. We choose the
graphs centered on the nodes that satisfy the
most constraints, presented here in the order of
their importance:
 w and wi must have the same part of
speech.
 w and wi have the same syntactic proper-
ties. If w and wi are verbs, they must have
the same subcategorization structure.
 w and wi are the same lemma. We empha-
size that a graph centered on a different
lemma, but with the same subcategoriza-
tion structure is preferred to a graph with
the same lemma, but a different subcate-
gorization structure.
 The edge representing the word pair to which
we want to assign a semantic relation has a
match in G(wi). From all graphs that com-
ply with this constraint, the ones that have the
lowest distance  corresponding to the high-
est matching score  are chosen. The graphs
are matched edge by edge. Two edges match
if the grammatical relation and the connectives
match. Figure 1 shows the formula that com-
putes the distance between two graphs. We
note that edge matching uses only edge infor-
mation  grammatical and connective informa-
tion. Using the node information as is (lemmas
and their part of speech) is too restrictive. We
are looking into using word similarity as a so-
lution of node matching.
If no matching graph has been found, the system
searches for a simpler match, in which the current
word pair is matched against previously processed
pairs, using the same formula as for edge distance,
and preferring the pairs that have the same modi er.
This algorithm will retrieve the set of graphs
fG(wi)g, which give the same score when matched
84
De nition of a graph centered on w:
G(w) = fwi,edge(w,wi) or edge(wi,w)jwiappears in sent. S, and is connected to wg
edge(w,wi) = fGRi,Conig ; GRi 2 fsubject,object,complement,...g
Coni 2 fat,in,on,with,for,...g
Distance metric between two graphs:
dist(G(w1),G(w2)) = Nsummationtext
k=1
d(edge1k,edge2k); edgeik 2 G(wi),N is the number of edges in G(wi)
d(edge1k,edge2k)=d(fGR1k,Con1kg,fGR2k,Con2kg)
=d1(GR1k,GR2k) + d1(Con1k,Con2k)
d1(x,y) =
braceleftbigg 0 : x = y
1 : x 6= y
Figure 1: Distance between two graphs
with the current graph. The set of possible semantic
relations presented to the user consists of the seman-
tic relation on the edge of each G(wi) that matches
the edge (of the current graph) corresponding to the
word pair which we are analysing.
To the sentence:
When you breathe out on a cold day, you make a
cloud.
corresponds the following syntactic graph:
(compl,nil)
(subord,when)out
(v, sv)
(compl,nil)
(v,svo)breathe make
cold
day
you you
(subj,nil)
(subj,nil)
(compl,on) cloud
When we focus on the graph centered on a spe-
ci c word, such as breathe, we look only at the node
corresponding to the word breathe, and the nodes
adjacent to it.
To process a pair P = (wH,wM), the system  rst
builds G(wH), and then searches through previously
stored graphs for those which have the same center
wH, or have the same part of speech as wH assigned
to its central node. For each graph found, we com-
pute a distance that gives a measure of the match
between the two graphs. The best match will have
the smallest distance.
For example, for the sentence:
Weathermen watch the clouds day and night.
the system builds the following network centered
on the predicate watch2:
cloudweatherman
watch
(v, svo)
(subj,nil)
(compl,nil)
day and night
(compl,nil)
The system locates among previously stored net-
works those centered around verbs3. For the sen-
tence above, the system uses the following graph,
2The nil value on the edges means that no preposition or
other connective explicitly links the two words or the corre-
sponding syntactic structures.
3If more detailed information is available, the system will
choose only networks associated with verbs that have the same
subcategorisation pattern (svo, svoi and so on).
85
built from the immediately preceding sentence in the
text:
Air pilots know that clouds can bring rain, hail,
sleet and snow.
(v, svo)
(subj,nil)
know
air pilots bring
(compl,nil)
AGENT
OBJECT
According to the metric, the networks match and
the pairs (watch, weatherman) and (know, air pi-
lots) match, so the semantic relation for the pair
(know, air pilots) is proposed as a possible relation
for pair (watch, weatherman) .
5 Experiments
The system processes the 513 sentences interac-
tively. It begins by running the DIPETT parser.
Next, it extracts syntactic units (clauses, phrases,
words) and pairs them up according to the informa-
tion in the parse tree. Each unit is represented by
its head word. Next, the system checks if the same
pair of word lemmas has already been processed, to
propose the same relation(s) to the user as options.
If not, the system builds a graph centered on the
head word, and proceeds with the matching on pre-
viously encountered instances, as described in sec-
tion 4. When a set of candidates has been found, the
system goes through a dialogue with the user.
The system generated 2020 pairs from the 513
sentences. The experiment was run in 5 interactive
sessions of approximately 3 hours each. The total
net processing time was 6 hours, 42 minutes and 52
seconds4. While it would have been instructive to
run the system several times with different users, it
was not feasible. The experiment was run once, with
two cooperating users. They were instructed on the
set of semantic relations, and told how the system
works. They discussed the semantic relation assign-
ment and, once agreed, compared the system’s sug-
gestion with their decision.
DIPETT did not produce a complete parse for all
sentences. When a complete parse (correct or incor-
rect) was not possible, DIPETT produced fragmen-
tary parses. The semantic analyser extracted units
even from tree fragments, although sometimes the
fragments were too small to accommodate any pairs.
Of the 513 input sentences, 441 had a parse tree that
allowed the system to extract pairs.
4The time difference accounts for system processing times,
and user interaction for other steps of the analysis.
# of analyzed examples 1475
level statistics CL IC NP
64 978 433
user actions accept choose supply
459 393 623
avg. # of suggestions 2.81
graph-matching usage 933
level/action statistics CL IC NP
accept 183 (19.61%) 9 141 33
choose 349 (37.41%) 23 314 12
supply 401 (42.98%) 27 316 58
Table 1: Summary of semantic analysis
Of 2020 pairs generated, the users discarded 545
in the dialogue step. An example of an erroneous
pair comes from the sentence:
Tiny clouds drift across like feathers on parade.
The semantic analyser produced the pair
(drift,parade), because of a parsing error: pa-
rade was parsed as a complement of drift, instead
of a post-modi er for feathers. The correct pairing
(feather,parade) is missing, because it cannot be
inferred from the parse tree.
Table 1 gives a summary of the processing statis-
tics. We observe that graph-matching was used to
process a clear majority of the total pairs extracted  
63.25% (933/1475) , leaving the remaining 36.75%
for the other two heuristics and for cases where no
suggestion could be made. In 57.02% of the situa-
tions when graph-matching was used, the system’s
suggestion contained the correct answer (user’s ac-
tion was either accept or choose), and in 19.61% of
the situations a single correct semantic relation was
proposed (user action was accept).
When the system presents multiple suggestions to
the user, including the correct one, the average num-
ber of suggestions is 3.75. The small number of
suggestions shows that the system does not simply
add to the list relations that it has previously encoun-
tered, but it learns from past experience and graph-
matching helps it make good selections. Figure 2
plots the difference between the number of exam-
ples for which the system gives the correct answer
(possibly among other suggestions) and the number
of examples when the user must supply the correct
relation, from the  rst example processed until the
end of the experiment. We observe a steady increase
in the number of correctly processed examples.
Our system does not differentiate between syntac-
tic levels, but based on the structures of the syntac-
tic units in each pair we can decide which syntactic
level it pertains to. For a more in-depth analysis, we
have separated the results for each syntactic level,
86
-20
 0
 20
 40
 60
 80
 100
 120
 140
 0  100  200  300  400  500  600  700  800  900  1000
Difference in number of examples
Examples processed
Figure 2: Difference between the number of situa-
tions in which the user accepts or chooses from the
system’s suggestions, and when it must supply the
correct relation
and present them for comparison in Figure 3.
We observe that the intra-clause level  verbs
and their arguments  makes the best use of graph-
matching, with the curve showing the cumulative
number of situations in which the system makes cor-
rect predictions becoming steeper as more text is
processed. At the same time, the curve that plots the
cumulative number of cases in which the user has to
supply a correct answer begins to level off. As ex-
pected, at the noun-phrase level where the syntactic
structures are very simple, often consisting of only
the noun and its modi er (without even a connec-
tive), the graph-matching algorithm does not help as
much. At the inter-clause level the heuristic helps,
as shown by the marginally higher curve for cumula-
tive accept/choose user actions, compared to supply
actions.
6 Conclusions
We have shown through the results gathered from an
interactive and incremental text processing system
that syntactic-semantic graph-matching can be used
with good results for semantic analysis of texts. The
graph-matching heuristic clearly dominates other
heuristics used, and it learns to make better predic-
tions as more examples accumulate.
Graph-matching is most useful for assigning se-
mantic relations between verbs and their arguments,
but it also gives good results for inter-clause rela-
tions. At the noun-phrase level, we could only tackle
noun-modi er pairs that exhibit a modicum of syn-
tactic structure  a connective. For base NPs there
is practically nothing that syntactic information can
bring to the semantic analysis process.
The graph-matching process could be improved
by bringing into play freely available lexical re-
1. All syntactic levels
 0
 100
 200
 300
 400
 500
 600
 0  100  200  300  400  500  600  700  800  900  1000
Cummulative number of examples
Examples processed
accept+choose
supply
2. Clause level (CL)
 0
 5
 10
 15
 20
 25
 30
 35
 0  10  20  30  40  50  60
Cummulative number of examples
Examples processed
accept+choose
supply
3. Intra-clause level (IC)
 0
 50
 100
 150
 200
 250
 300
 350
 400
 450
 500
 0  100  200  300  400  500  600  700  800
Cummulative number of examples
Examples processed
accept+choose
supply
4. Noun phrase level
(NP)
 0
 10
 20
 30
 40
 50
 60
 0  20  40  60  80  100  120
Cummulative number of examples
Examples processed
accept+choose
supply
Figure 3: Graph-matching for different syntactic
levels
87
sources. For now, the actual words in the graph
nodes are not used at all. We could use WordNet
to compute word similarities, to select closer match-
ing graphs. VerbNet or FrameNet information could
help choose graphs centered on verbs with simi-
lar syntactic behaviour, as captured by Levin’s verb
groups (Levin, 1993) which are the basis of VerbNet.

References
Jordi Atserias, L. Padr·o, and German Rigau. 2001. Integrating
multiple knowledge sources for robust semantic parsing. In
Proceedings of RANLP - 2001, Tsigov Czark, Bulgaria.
Collin F. Baker, Charles J. Fillmore, and John B. Lowe. 1998.
The Berkeley FrameNet project. In COLING-ACL, pages
86 90, Montreal, Canada.
Ken Barker, Terry Copeck, Sylvain Delisle, and Stan Szpakow-
icz. 1997a. Systematic construction of a versatile case sys-
tem. Journal of Natural Language Engineering, 3(4):279 
315.
Ken Barker, Sylvain Delisle, and Stan Szpakowicz. 1997b.
Test-driving TANKA: Evaluating a semi-automatic system
of text analysis for knowledge acquisition. In Proceedings
of CAI 1997, pages 60 71, Vancouver, BC, Canada.
Xavier Carreras and Lluis Marquez, editors. 2004. Introduction
to the CoNLL-2004 Shared Task: Semantic Role Labelling.
Boston, MA, USA.
Xavier Carreras and Lluis Marquez, editors. 2005. Introduction
to the CoNLL-2005 Shared Task: Semantic Role Labelling.
Ann Arbour, MI, USA.
Noam Chomsky. 1970. Remarks on nominalizations. In Rod-
erick Jacobs and Peter Rosenbaum, editors, Readings in En-
glish Transformational Grammar, pages 184 221. Ginn and
Co., Waltham, MA, USA.
Peter Clark and Bruce Porter. 1997. Building concept reprezen-
tations from reusable components. In AAAI, pages 367 376,
Providence, Rhode Island.
Sylvain Delisle and Stan Szpakowicz. 1995. Realistic pars-
ing: Practical solutions of dif cult problems. In PACLING,
Brisbane, Queensland, Australia.
Sylvain Delisle, Terry Copeck, Stan Szpakowicz, and Ken
Barker. 1993. Pattern matching for case analysis: A com-
putational de nition of closeness. In ICCI, pages 310 315,
Sudbury, ON, Canada.
Charles Fillmore and Beryl T. Atkins. 1998. FrameNet and
lexicographic relevance. In Proceedings of the 1st Interna-
tional Conference on Language Resources and Evaluation,
Granada, Spain.
Charles Fillmore. 1968. The case for case. In Emmond Bach
and Robert T. Harms, editors, Universals in Linguistic The-
ory, pages 1 88. Holt, Rinehart and Winston.
Daniel Gildea and Daniel Jurafsky. 2002. Automatic labeling
of semantic roles. Computational Linguistics, 28(3):245 
288.
Fernando Gomez. 1998. A representation of complex events
and processes for the acquisition of knowledge from text.
Kowledge-Based Systems, 10(4):237 251.
Jeffrey Gruber. 1965. Studies in Lexical Relations. Ph.D.
thesis, MIT, Cambridge, MA. Reprinted in Jeffrey Gru-
ber. 1976. Lexical Structures in Syntax and Semantics. Part
I. North-Holland Publishing Company, Amsterdam.
Richard D. Hull and Fernando Gomez. 1996. Semantic inter-
pretation of nominalizations. In 13th National Conference
on Arti cial Intelligence, pages 1062 1068, Portland, Ore-
gon, USA.
Karin Kipper, Hoa Trang Dang, and Martha Palmer. 2000.
Class-based construction of a verb lexicon. In AAAI/IAAI,
pages 691 696.
Nancy Larrick. 1961. Junior Science Book of Rain, Hail, Sleet
and Snow. Garrard Publishing Company, Champaign, Illi-
nois.
Beth Levin. 1993. English Verb Classes and Alternations. Uni-
versity of Chicago Press.
Vidya Niwas Misra. 1966. The Descriptive Technique of
Panini. Mouton, The Hague.
Sameer Pradhan, Wayne Ward, Kadri Hacioglu, James H. Mar-
tin, and Daniel Jurafsky. 2005. Semantic role labelling us-
ing different syntactic views. In Proceedings of ACL 2005,
pages 581 588, Ann Arbour, MI, USA.
Randolph Quirk, Sidney Greenbaum, Geoffrey Leech, and Jan
Svartvik. 1985. A Comprehensive Grammar of the English
Language. Longman, London and New York.
Barbara Rosario and Marti Hearst. 2001. Classifying the se-
mantic relations in noun-compounds via a domain speci c
hierarchy. In EMNLP, pages 82 90, Pittsburg, PA, USA.
Barbara Rosario, Marti Hearst, and Charles Fillmore. 2002.
The descent of hierarchy and selection in relational seman-
tics. In ACL, Philadelphia, PA, USA.
Lei Shi and Rada Mihalcea. 2005. Putting pieces together:
Combining framenet, verbnet and wordnet for robust seman-
tic parsing. In Proceedings of CICLing 2005, pages 100 
111, Mexico City, Mexico.
Lucien Tesni ere. 1959. ·El·ements de syntaxe structurale. C.
Klincksieck, Paris.
