Semantic Processing for Text Entailment with VENSES

Rodolfo Delmonte, Sara Tonelli, Rocco Tripodi

Department of Language Sciences
Università Ca’ Foscari – Ca’ Bembo
30123, Venezia, Italy
[email protected]

Abstract valid meaning correspondence. VENSES has a

discourse level anaphora resolution module,
In this paper we present two new mechanisms we
coupled to a full-fledged semantic interpreter,
created in VENSES, the system for semantic
which is paramount to allow entailment in pairs
evaluation of the University of Venice. The first
where the relevant portion of text contains
mechanism is used to match predicate-argument
pronominal expressions.
structures with different governors, a verb and a
The system is organized into twelve layers as
noun, respectively in the Hypothesis and the Text.
described below – but see [1,3,5] for a complete
It can be defined Augmented Finite State
description:
Automata (FSA) which are matching procedures
• Tokenizer and sentence splitting;
based on tagged words in one case, and
• Tagger from dictionary lookup or from
dependency relations in another. In both cases, a
morphological analysis;
number of inferences – the augmentation - is fired
• Tag disambiguation with finite-state automata
to match different words. The second mechanism
and the aid of lexical information;
is based on the output of our module for anaphora
• Head-based Chunk building phase;
resolution. Our system produces antecedents for
• Recursive argument/adjunct (A/A) constituent
pronominal expressions and equal nominal
building procedure as a list of syntactic-semantic
expressions. On the contrary, no decision is taken
structures with tentative GFs labels;
for “bridging” expressions. So the “bridging”
• Clause builder that takes as input the A/A vector
mechanism is activated by the Semantic Evaluator
and tries to split it into separate clauses;
and has access to the History List and the semantic
• Recursive clause-level interpretation procedure,
features associated to each referring expression. If
that filters displaced or discontinuous constituents;
constraint conditions meet, the system looks for a
• Complex sentence organizer which outputs DAG
similar association of property/entity in web
structures;
ontologies like Umbel, Yago and DBPedia. The
• Logical Form with syntactic indices and
two mechanisms have been proven to contribute a
Semantic Roles;
5% and 3% accuracy, respectively.
• Transducer from DAGs to AHDSs by recursive
calls;
1. Introduction
• Pronominal Binding at clause level followed by
Anaphora Resolution at intersentential level;
Venses, the system for semantic processing,
• Semantic Module builds propositional level
represents a linguistically-based approach for
feature vectors, which also contain discourse
semantic inference [2] which is built around a neat
relations.
division of labour between two main components.
Text entailment in VENSES is interpreted and
The first is a grammatically-driven subsystem
implemented in four different steps:
which is responsible for the level of predicate-
- semantic similarity: in general terms by
arguments well-formedness and works on the
searching entities and events of the T/H pairs and
output of a deep parser that produces augmented
comparing them on the basis of dictionaries,
head-dependency structures. A second subsystem
computational lexica and ontologies;
fires allowed logical and lexical inferences on the
- propositional level constraints: this is what we
basis of different types of structural
also call General Consistency checks which are
transformations intended to produce a semantically
targeted to high level semantic attributes like
presence of modality, negation, and opacity a mechanism for Centering. The system looks for
operators, the latter ones are expressed either by antecedents of pronominal expressions, and
the presence of discourse markers of conditionality anchors for bridging coreference of non identical
or by a secondary level relation intervening nominal expressions. The mechanism
between the main predicate and a governing higher implemented tries to bridge the gap between
predicate belonging to the class of non factual linguistic knowledge and commonsense
verbs. The governing head predicate is responsible knowledge or knowledge of the world, represented
for the factitivity of the dependent. Opacity is here as UMBEL, the web ontology.
determined by type of governing predicates, The second one – discussed in this section - uses
basically those belonging to the class of nonfactive Augmented FSA for nominal paraphrases, i.e. it
predicates. Modality is revealed by the presence of searches for nominal paraphrases of the copulative
modal verbs at this level of computation. Modality construction contained in the Hypothesis. In some
could also be instantiated at sentence level by cases the construction may be governed by what
adverbials, and be verified by General Consistency we define as a “light verb” construction, that is a
Checks. Finally, negation may be expressed verb that may be paraphrased by a preposition in a
locally as adjunct of the verb, but also as negative nominal construction: some such verbs are “locate,
conjunction and negative adverbial. It may also be situate” which may be represented as “in” in the
present in the determiner of the nominal head and corresponding NP.
checked separately when comparing referring The Augmented FSA are organized either as
expressions considered in the inference. Negation matching procedures based on tagged words, or on
may also be lexically incorporated in the verb; dependency structures. In both cases, additional
- if general constraints are satisfied, we look for inferential processes are called for in order to
argument and adjuncts correspondences and for match non identical linguistic descriptions. Here
their possible paraphrases, taking care of below are some example to illustrate the two
inadmissible cases on the basis of semantic roles; techniques.
- inside arguments and adjuncts, we look for
quantifiers and generic numerical specification by 2.1. Using Augmented FSA for SE with
searching modifiers of nominal heads; we also Nominal Paraphrases
look for non intersective modifiers;
- at adjunct level we take care of spatio-temporal The procedure takes as input the tagged list of
modifiers if any and check for their semantics. words making up the Hypothesis and tries to
The problem posed by this year text level RTE match with the relevant portion of the Text that
which needs to be addressed fully requires a full- contains similar words and tags, as follows
fledged system for anaphora resolution both at
pronominal and nominal (bridging) level (see [4] match_template(Hypothesis, Text)
for a complete description. What we mean by a
full-fledged system will be clarified here below: If the match succeeds the semantic evaluation
suffice it to say that in order for bridging to apply outputs a value that is indicative of the type of
we regard it mandatory that anaphora resolution be decision taken. This matching produce is reached
enacted by a Topic Hierarchy mechanism in the by the analysis only after General Consistency
vein of what Grosz and Sidner suggested in their Checks have passed at higher propositional level.
papers [7,8]. Consider the first example where we highlight the
portions of T/H relevant for the semantic
2. Two New Mechanisms at Work in Getaruns evaluation:

RTE5 – T/H Pair 364 – Entailment=True

In this year version of the Semantic Evaluator we Trains, trams, cars and buses ground to a halt on
implemented two new mechanisms. The first one – Monday after a shoot-out between 18:00 CET and
discussed in the next section – uses the output of 19:00 CET in the historical city of Basel in
our system for anaphora resolution which is based Switzerland. The first shots were fired at around
16:00 CET according to the spokesperson of the
on Sidner and Grosz’s intuition on the importance Government. Special Police Forces were getting
of the presence of a Topic Hierarchy in addition to prepared to launch an assault against this building.
 The authors of the shoot-out are unknown as well
as their target. Railway traffic was blocked off at RTE3 – T/H Pair 173 – Entailment=True
around 18:00 CET, according to the Swiss Federal Prince Laurent of Belgium, the youngest
Railways’ ( SBB-CFF-FFS ) spokesperson, son of King Albert II of Belgium, has
Christian Krìuchi. At around 19:00 CET, railway been questioned last night by the federal
traffic resumed . police and is attending today's court
Basel is a European city. session in Hasselt in a marine fraud case
that has gripped Belgian media since last
In more detail, Augmented Finite State Automata December.
King Albert II of Belgium is the father
mean that in addition to equality matching that is of Prince Laurent.
at the basis of the whole algorithm, the system
looks for inferences and other lexical information Here below we list the augmented dependency
to authorize the match. In fact, these procedures as relations for the text and hypothesis where the
a whole allow the matching to become more indices at the end of each term are taken from the
general with constraints though. The instructions corresponding syntactic constituent and identify
reported below are expressed in Prolog which uniquely in each sentence the semantic heads:
treats words constituted by or beginning with
upper case letters to be treated as variables. T. 'King_Albert' - ncmod / specif - 'Belgium' - sn6,
Constants on the contrary are written with lower 'King_Albert' - det - 'II' - sn5,
case letter, as for instance the words “of” and “in” 'Prince_Laurent' – ncmod/specif - son - sn1,
below. son - ncmod / specif - 'King_Albert' - sn5

match_template ([A,Is-_,T-_,F-_,G|Hyp],[G,of-_,A,in- H. 'King_Albert' - ncmod / specif - 'Belgium' - sn3,

_,L-_|Text]):- 'King_Albert' - det - 'II' - sn1,
lightsvbs(Is), father - ncmod / specif - 'Prince_Laurent' - sn4
high_rank(T,Lex), be - xcomp / prop - father - sn2,
locwn(L), 'King_Albert' - subj / theme_unaff - sn1
is_in(L,F1),
(natl(F1,F,_);natl(F1,_,F)), !. where we see that relations are reversed and need
to be checked carefully. In the text we know that
where the procedure “lightverbs” looks for Prince Laurent is the son of King Albert of
copulative verbs, i.e. the verb of the Hypothesis Belgium; on the contrary in the Hypothesis we are
must be a copulative verb; “high_rank” looks for told that King Albert of Belgium is the father of
high frequency words like articles; “locwn” Prince Laurent. Now, what is needed is the
verifies that the word present in the variable “L” is possibility to draw inferences about the
a location. Then there are two inferences: the first complementariness in the relation existing
one is fired by the call “is_in” that recovers the between father and son.
name of the continent in which “L” belongs, thus The code related to the semantic evaluator is
implicitly requiring “L” to be a name of a nation. shown here below:
Then the second inference looks for the
corresponding nationality adjective. Values for the searchmatchtemprels(Text_rels,Hypos_rels):-
variables are then as follows: remove(be-xcomp/prop-R-Sn, Hypos_rels,Rests),
L --> Switzerland relatives(Rels),
on(R,Rels),
F1 --> Europe
remove(R-xcomp/prop-Sn2,Rests,Resto),
F --> European remove(Head-subj/Rol-Sn1,Resto,Rest),
A = Basel-np, Is = is, T = a, G = city-n remove(R1-ncmod/specif-Head1-Sn3,
Text_rels,Rest1),
2.2. Using Dependency Relations for SE with on(R1,Rels),
Nominal Paraphrases Head=Head1,
(compl_rel(R1,R);
In the second mechanism we activated, FSA are compl_rel(R,R1)),
used to match dependency relations. Consider the …
second example:
compl_rel(father,son). In order to know precisely where the pope actually
compl_rel(father,daughter). landed with his helicopter, the adjective
… “homeless” is needed. But to recover this piece of
information, the system has to be able to bridge
The procedure looks for a copulative construction the information expressed in the first sentence to
in the Hypothesis and then checks to see whether the information of the second one. This can only
the property predicated is one of the set of be done with a system of anaphora resolution that
“relatives”. If yes, it removes the property and the can compute the bridging relation [9,10]
predicated subject of the property to use it for intervening between the first occurrence of CAMP
matching purposes with the corresponding as an indefinite expression, and the second
relations in the Text. Then it removes modifier occurrence in second sentence as definite
relations of specification in the Text pool of expression. This will be explained in more details
relations where the governing head is identical to below.
the subject head of the Hypothesis and the One such system is shown in Fig. 1 below, where
property head is in a complementary relation with we highlight the architecture and main processes
the corresponding head found in the Text. The undergoing at the anaphora level. First of all, the
following portion of the procedure will recursively subdivision of the system into two levels: Clause
eliminate all identical relations until the level – intrasentential pronominal phenomena –
Hypothesis pool is empty. where all pronominal expressions contained in
modifiers, adjuncts or complement clauses receive
3. Topic Hierarchy And Bridging their antecedent locally. Possessive pronouns,
pronouns contained in relative clauses and
A conspicuous number of T/H pairs of this year complement clauses choose preferentially their
datasets is characterized by the need to look for antecedents from list of higher level referring
semantic attributes and properties asserted in the expressions. Not so for those pronouns contained
Hypothesis in different sentences contained in the in matrix clauses. In particular the ones in subject
Text. So that the possibility to match T/H one position are to be coreferred in the discourse. This
sentence at a time and produce a sufficient requires the system to be equipped with a History
environment for the semantic evaluation is simply List of all referring expressions to be used when
no longer available. One such case is represented needed.
by Intersective adjectives and is presented here
below:

RTE5 – TH Pair 81 - Entailment=True

The pope flew by helicopter from Vatican
City to a tent camp near the village of Onna,
where he led a prayer in the cold and rainy
weather for the hundreds killed in the April 6
quake. The camp houses hundreds of families
left homeless when the magnitude-6.3 quake
destroyed their homes. Residents there
welcomed the pope, who kissed and hugged
some of the children. "I have come here
personally to this splendid and hurt land of
yours, which is living days of great pain and
precariousness, to express in the most direct
way my kind closeness", the pope told
residents. "I've followed the news with
apprehension, sharing with you your
consternation... for the dead, along with your
anxious worries about how much you've lost in
a brief moment.
The pope reached the homeless camp near
Onna by helicopter.
Fig. 1 Anaphoric Processes in VENSES
In the system, three levels are indicated: Clause don’t find in the Text – that is and the semantic
level, i.e. simple sentences; Utterance level, i.e. evaluation based on predicate-argument structures
complex sentences; Discourse level, i.e. has failed - we switch to the Bridging coreference
intersententially. Our system computes semantic module. The Hypothesis will typically contain two
structures in a sentence by sentence fashion and linguistic descriptions associated by means of
any information useful to carry out anaphoric some “light verb”. The module will look for one of
processes needs to be made available to the the nominal expressions used in the Hypothesis
following portion of text, and eventually to the either in definite, indefinite or proper name
Semantic Evaluation that computes entailment. As linguistic description, in a sentence that precedes
a first case, I will present an example of the one containing the other linguistic descriptions
pronominal anaphora resolution which requires the of the Hypothesis. This procedure starts by
system to identify genre and number in order to searching in the History List where referring
reject local possible antecedents. The pronoun we expressions are listed by sentence number and
are referring to is the first HER highlighted in the their rhetorical label in the Topic Hierarchy – i.e.
text below: they can be computed as either Main Topic,
Secondary Topic or Potential Topic [6,7,8]. We
RTE5–TestSet - TH Pair 42 - search for Potential or Main Topics because the
Entailment=True definite description would constitute a new
DENVER - Angie Zapata was a tall woman
with striking black hair and eyes who would linguistic item in the History List; on the contrary
attract the attention of men, even those who Secondary Topic would be used only in case the
knew she was biologically male. But prosecutors linguistic description has already been asserted as
say when Allen Andrade found out, he beat her Main Topic in previous stretch of discourse and is
to death with a fire extinguisher. Her sister
discovered her battered body under a blanket in
accompanied by another Topic.
her Greeley apartment last July. Andrade, 32, of The mechanism implemented tries to bridge the
Thorton, is scheduled to go on trial Tuesday on gap between linguistic knowledge and
charges including first-degree murder and a bias- commonsense knowledge or knowledge of the
motivated crime, which could add three years to world, represented here as web ontologies like
his prison sentence if convicted.
Angie Zapata has been killed with a fire UMBEL, YAGO and DBPEDIA [11]. In fact,
extinguisher. what we do is looking for external knowledge
whenever our linguistic procedures require it. The
Of course the system has to be able to deal with reason for this move is motivated as follows:
feminine Names as opposed to masculine ones. - linguistic descriptions to be matched and
Then there must be a History List and a Centering related by a bridging coreference link are
mechanisms. Both are needed in order to choose DIFFERENT
the appropriate antecedent for HE and HER. - the usual semantic relations made
Obviously syntactic structure needs to be present available in WordNet and also by the
in order to be able to assign Allen_Andrade as lexical fields of Moby have failed
possible antecedent of HE: the pronoun and the - we don’t have enough confidence due to
proper name are positioned in different clauses and the presence of other possible coreferents
do not undergo have to obey to disjointness In other words, this strategy is used to restrain,
constraints. Eventually, the Semantic Evaluator reduce, limit access to external knowledge of the
has to be able to compute passive structures and world to those cases that really require it. Thus
compare it to active structures, and this is done by eliminating all T/H pairs which are lacking such
looking at Semantic Roles and Grammatical enabling conditions and must be regarded as
Functions. Then the last problem to be solved is FALSE cases.
the semantic similarity of KILL with We will comment a number of significant
BEAT_TO_DEATH that must be defined examples to clarify the way in which our system
somewhere if the entailment has to hold. operates.
As commented above, the idea is that every time
we have a predication in the Hypothesis that 3.1. Topic Hierarchy And Bridging
associated a property to a proper noun that we
We chose three examples from the two sets definiteness and/or semantic type of the head
Development and Test, and indicated in a headline noun. These information are in our encoded in the
the configuration of Bridging Expression and its vector associated to each referring expression
preceding Coreferent. semantic head, as follows:

Proper Name + Definite Expression ref_ex(sn1, Madonna, [+ref, def0, nil, nil, -pro, -ana, -
Rte5 – DevSet - TH Pair 10 class], 3, fem, sing, [human], subj/theme_unaff)
ref_ex(sn1, existence, [+ref, +def, very, nil, -pro, -ana,
(CNN) -- Malawians are rallying behind Madonna +class], 3, neut, sing, [place, state], subj/actor)
as she awaits a ruling Friday on whether she can ref_ex(sn9, it, [+ref, +def, nil, nil, +pro, -ana, +le], 3, neu,
adopt a girl from the southern African nation. The sing, [any], subj/agent)
pop star, who has three children, adopted a son
from Malawi in 2006. She is seeking to adopt As can be seen from representations, proper nouns
Chifundo “Mercy” James, 4. “Ninety-nine percent are marked def0, +ref and –class; on the contrary
of the people calling in are saying, let her take the
baby,” said Marilyn Segula, a presenter at Capital common nouns are marked +def/-def, +ref and
FM, which broadcasts in at least five cities, +class. Pronouns do not have the attribute CLASS
including the capital, Lilongwe. but +/-le which stands for Lexically Expressed.
Madonna has three children. The vector includes Functional Features – Person,
Gender, Number – and Semantic Features in the
Indefinite Expression + Proper Name
sense of General Nouns or Inherent Features. At
Rte5 – TestSet - TH Pair 83
the end of the vector or Prolog term, we report
A Ugandan spy who set up a bogus charity and grammatical function and semantic role associated
embezzled thousands of dollars of funding meant to the head noun which can be found in syntactic
for Aids patients has been jailed for 10 years. and dependency representations by means of the
Teddy Sseezi Cheeye, 51, took $56,000 (£38,000) index, positioned at the beginning.
from the Global Fund charity, which aims to
prevent HIV, tuberculosis and malaria. He set up
an NGO, the Uganda Centre for Accountability, 4. Evaluation and Ablation test
which received cash in 2005 to do HIV/Aids
community work. But the High Court in Kampala The evaluation results we present try to give a
heard Cheeye siphoned off the funds instead.
Teddy Sseezi Cheeye is an Ugandan spy.
comprehensive picture of the system performance
over the overall datasets made available with RTE.
Definite Expression + Proper Name It is worthwhile reminding that the first 2
Rte5 – TestSet - TH Pair 269 challenges contained very short Texts if compared
to the Text average size of the following
The eruption happened at around 1:30 PM local challenges. In particular then, Texts contained in
time, the United States Geological Survey RTE4 testset and RTE5 development and test sets
reported. The volcano had erupted four times on
Friday, billowing ash up to 51,000 feet up into the are much longer then those contained in RTE3.
air. These are the latest in a series of eruptions The difference in treatment of these datasets is
from Mount Redoubt, which started on March 22. quite obvious: modeling a paragraph long text is
The volcano had not erupted since a four-month certainly much harder. In addition, RTE5 texts
period in 1989-90. The Alaska Volcano
Observatory set its alert level at red, the highest
have a certain number of T/H pairs where the
possible level, meaning that an eruption is contents of the Hypothesis is scattered amongst a
imminent, and that it would send a "significant number of sentences in the paragraph. This makes
emission of volcanic ash into the atmosphere." the task much harder than in all those cases in
Mount Redoubt is located in Alaska. which semantic matching can be concentrated on
just one sentence in the Text paragraph.
As can be gathered from the headers and the
As will be noticed in the data reported in Table 1.
highlighted portions of texts, the cases to be
below, there is a remarkable difference in the
covered all involve a proper noun which can be
results obtained in the Development and the Test
either a person’s name or a location. There are
set: 10 percent point accuracy. A possible reason
three different configurations to account for, which
for this is the fact that RTE5 Testset contains a lot
require basically a search for the type of
more cases of difficult to spot entailment relations.
It is a fact, that a great number of T/H pairs interesting to notice that not all task behave in the
contain Texts where the relevant relations are same way.
scattered in more than one sentence, thus making This type of “sloppy” semantic similarity
the semantic matching task harder to perform. matching is fired every time the system needs
approximated or fuzzy similarity information. In
Subtask Accuracy(%) particular, it is never permitted whenever precise
ir 61.00 information is required, as for instance in what we
qa 65.00 call General Consistency Checking procedures.
ie 58.50 These procedures are carried out to check for the
Average 61.50
presence of Quantified Expressions, information
Table 1: Official results for Run 1 of our system – No
Ranking
related to Spatio-Temporal Location, and any kind
Subtask Precision(%) of numerical information present in the Hypothesis
ir 58.62 that has to be present also in the Text. On the
qa 67.14 contrary, whenever we look for attributes,
ie 66.00 modifiers and other similar adjuncts of the
Average 64.45 arguments expressed in predicate-argument
Table 2: Official results for Run 2 of our system – Ranking structure, we allow access to lexical fields
contained in thesaura. This may also apply in all
Results for past RTE datasets as a whole fare on copulative constructions, whenever a certain
average 63% - but see table 4. below. Results for property is being associated to the subject of the
the Contradiction Dataset, are as follows: predication.
Accuracy measured as ratio of Correct Pairs/All These matching procedure are scattered all over
Pairs: 108/131 = 0.8245 the evaluation algorithm: what we did was simply
Results for the Development and the Test set of dummifying the access to the matching
RTE 5, are as follows: procedures, by inserting a dummy couple of values
DEVELOPMENT set: Accuracy measured as ratio – nil, nil – in place of the two variables that had to
of Correct Pairs/All Pairs: 0.73 be taken into consideration by the matching
It is important to notice that in all cases with no procedure, and inserted a cut – in Prolog an
exception whatsoever, the percentage of True T/H instruction not to allow recursion and oblige a
pairs found is higher than the percentage of False. failure – in place of the procedure itself, which
was hidden.
4.1 Ablation Test
Subtask Accuracy1 Accuracy2
We carried out one ablation test where we ir 61.00 65.50
removed matching procedures related to Grady qa 65.00 58.00
Ward’s MOBY Thesaurus as well as to Roget’s ie 58.50 52.50
Thesaurus. In fact what we eliminated was a Average 61.50 58.67
procedure which used “lexical fields” as semantic Table 3: Ablation Test results compared with Run1
similarity matching in all cases of non identical
lemmas. We used this procedure after eliminating
cases of antonymy which could degrade the 5. Conclusions and Future Work
semantic similarity matching. After the filter for
antonyms, matching was carried out on lemmas as We presented our improvements to VENSES, our
usual. Access to Thesaura can in some cases system for semantic evaluation, which uses a
contribute important and relevant information, but proprietor complete system of text analysis based
this is not always guaranteed as shown by the on a deep system called GETARUNS. We
results of the test reported here below. In introduced a number of new modules that take
particular, we may notice that in one case, IR advantage of the output of the anaphora resolution
subtask, we improved accuracy by 0.045 points. algorithm and exploit its representations to attempt
So, even though in the remaining subtasks there is bridging coreference. In case constraints are
always a reduction of the overall accuracy, it is respected, the system looks for similar relations in
web ontologies, to confirm the anaphoric link. We
also implemented Augmented FSA both at tagging We report here below a table with the overall
and at dependency levels. The results are very results the system obtains on all RTE datasets.
encouraging and we saw an improvement of 8%
overall.

Table 4: VENSES overall results with RTE datasets

References

1. Delmonte R., 2009. Computational Linguistic Text

Processing – Lexicon, Grammar, Parsing and
Anaphora Resolution, Nova Science Publishers, 6. Delmonte, R., Bianchi D.: Binding Pronominals with
New York. an LFG Parser. In: Proceeding of the Second
2. Delmonte R., A. Bristot, M.A.Piccolino Boniforti, International Workshop on Parsing Technologies,
S.Tonelli (2007), Entailment and Anaphora Cancun (Mexico), ACL. (1991) 59-72
Resolution in RTE3, in Proc. ACL Workshop on 7. B. J. Grosz and C. L. Sidner. 1986. Attention,
Text Entailment and Paraphrasing, Prague, ACL intention, and the structure of discourse.
Madison, USA, pp. 48-53. Computational Linguistics, 12(3):175–204.
3. Delmonte R., 2007. Computational Linguistic Text 8. B. J. Grosz, A. K. Joshi, and S. Weinstein. 1995.
Processing – Logical Form, Semantic Interpretation, Centering. Computational Linguistics, 21(2):202–
Discourse Relations and Question Answering, Nova 225.
Science Publishers, New York. 9. K. Markert, M. Strube, and U. Hahn. 1996.
4. Delmonte R., A. Bristot, M. A. Piccolino Boniforti Inferential realization constraints on functional
and Sara Tonelli, 2006. Another Evaluation of anaphora in the centering model. In Proc. of 18th
Anaphora Resolution Algorithms and a Comparison Conf. of the Cog. Science Soc., pages 609–614.
with GETARUNS' Knowledge Rich Approach, 10. K. Markert, M. Nissim, and N.. Modjeska. 2003.
ROMAND 2006, 11th EACL, Trento, ACL, 3-10. Using the Web for nominal anaphora resolution. In
5. Delmonte R.(2002), GETARUN PARSER - A parser Proc. of the EACL Workshop on the Computational
equipped with Quantifier Raising and Anaphoric Treatment of Anaphora, pages 39–46.
Binding based on LFG, Proc. LFG2002 Conference, 11. N. Modjeska, K. Markert, and M. Nissim. 2003.
Athens, at http://cslipublications.stanford.edu/hand/ Using the Web in ML for anaphora resolution. In
miscpubsonline.html, 130-153. Proc. of EMNLP-03, pages 176–183.