Ontological resources and question answering 
 
Roberto Basili (*), Dorte H. Hansen (**),Patrizia Paggio (**),   
Maria Teresa Pazienza (*), Fabio Massimo Zanzotto (*) 
 
 
(*) Dip. di Informatica Sistemi e Produzione 
University of Rome “Tor Vergata” 
{basili,pazienza,zanzotto} 
@info.uniroma2.it  
 
 
(**) Centre for Language Technol-
ogy 
University of Copenhagen 
{patrizia,dorte}@cst.dk 
 
 
Abstract 
This paper discusses the possibility of build-
ing an ontology-based question answering 
system in the context of the Semantic Web 
presenting a proof-of-concept system. The 
system is under development in the MOSES 
European Project.  
Introduction  
Question Answering (QA) systems (as QA track of 
the Text Retrieval Conference (TREC-QA) competi-
tions (Voorhees 2001)), are able both to understand 
questions in natural language and to produce answers in 
the form of selected paragraphs extracted from very 
large collections of text. Generally, they are open-
domain systems, and do not rely on specialised concep-
tual knowledge as they use a mixture of statistical tech-
niques and shallow linguistic analysis. Ontological 
Question Answering systems, e.g. (Woods et al. 1972, 
Zajac 2000) propose to attack the problem by means of 
an internal unambiguous knowledge representation. As 
any knowledge intensive application, ontological QA 
systems have as intrinsic limitation related to the small 
scale of the underlying syntactic-semantic models of 
natural language. 
 
While limitations are well-known, we are still ques-
tioning if any improvement has occurred  since the de-
velopment of the first ontological QA system LUNAR. 
Several important facts have emerged that could influ-
ence related research approaches: 
a0  a growing availability of lexical knowledge bases 
that model and structure words: WordNet (Miller 
1995) and EuroWordNet (Vossen 1998) among 
others; some open-domain QA systems have proven 
the usefulness of these resources, e.g. WordNet in 
the system described in (Harabagiu et al. 2001). 
a0  the vision of a Web populated by “ontologically” 
tagged documents which the semantic Web initia-
tive has promoted; in case this vision becomes a re-
ality, it will require a world-wide collaborative 
work for building interrelated “conceptualisations” 
of domain specific knowledge 
a0  the trend in building shallow, modular, and robust 
natural language processing systems (Abney 1996,  
Hobbs et al. 1996, Ait-Moktar&Chanod 1997, 
Basili&Zanzotto 2002) which is making them ap-
pealing in the context of ontological QA systems, 
both for text interpretation (Andreasen et al. 2002) 
and for database access (Popescu et al. 2003). 
 Given this background, we  are investigating a new 
approach to ontology-based QA in which users ask 
questions in natural language to knowledge bases of 
facts extracted from a federation of Web sites and or-
ganised in topic map repositories (Garshol 2003). Our 
approach is being investigated in the context of EU pro-
ject MOSES1, with the explicit objective of developing 
an ontology-based methodology to search, create, main-
tain and adapt semantically structured Web contents 
according to the vision of the Semantic Web. MOSES is 
taking advantage of expertise coming from several 
fields: software agent technology, NLP, graph theory 
                                                           
1 MOSES is a cooperative project under the 5th Frame-
work Programme. The project partners are FINSA Consult-
ing, MONDECA, Centre for Language Technology, 
University of Copenhagen, University of Roma Tre, Univer-
sity of Roma Tor Vergata and ParaBotS. 
and text mining. The test-bed chosen in the project is 
related to the development of an ontology-based knowl-
edge management system and an ontology-based search 
engine that will both accept questions and produce an-
swers in natural language for the Web sites of two 
European universities. The challenges of the project are:  
a0  building an ontological QA system;  
a0  developing a multilingual environment which im-
plies the ability to treat several languages, and, im-
portantly, several conceptualisations.    
 
In this paper, after briefly describing how the project 
is trying to comply with the semantic Web vision, we 
will focus on question processing, and in particular on 
the way in which NLP techniques and ontological 
knowledge interact in order to support questions to spe-
cific sites or to site federations.  
 
An ontology-based approach to question 
answering 
 
In our ontological QA system, both questions and 
domain knowledge are represented by the same onto-
logical language. It is foreseen to develop the QA sys-
tem in two steps. First a prototypical implementation is 
planned to answer questions related to the current 
“state-of-affairs” of the site to which the question is 
posed. In a second step, given a “federation” of sites 
within the same domain, we will  investigate whether 
and how an ontological approach could support QA 
across the sites. Answering a question can then be seen 
as a collaborative task between ontological nodes be-
longing to the same QA system. Since each node has its 
own version of the domain ontology, the task of passing 
a question from node to node may be reduced to a map-
ping task between (similar) conceptual representations. 
To make such an approach feasible, a number of diffi-
cult problems must still be solved. In this paper, we will 
provide details on how: 
a0  to build on existing ontologies and interface be-
tween them and language resources;  
a0  to interpret questions wrt the ontological language;  
a0  to model the mapping task for federated questions. 
 
Building on off-the-shelf semantic Web on-
tologies 
One of the results of the Semantic Web initiative 
will be the production of many interrelated domain-
specific ontologies that provide the formal language for 
describing the content of Web documents. In spite of the 
freedom allowed in the production of new conceptuali-
sations, it is reasonable to expect that a first knowledge 
representation jungle will leave room to a more orderly 
place where only the more appreciated conceptualisa-
tions have survived. This is a prerequisite for achieving 
interoperability among software agents. In view of this, 
and since publicly available non-toy ontology examples 
are already available, the effort of adapting an existing 
ontology to a specific application is both useful and 
possible. This experiment is being conducted in MOSES 
to treat the university domain.   
 
Ontologies for the Semantic Web are written in for-
mal languages (OWL, DAML+OIL, SHOE) that are 
generalisations/restrictions of Description Logics 
(Baader et al. 2003). TBox assertions describe concepts 
and relations. A typical entry for a concept is: 
 
ID Course 
Label Course 
Subclassof Work 
 
Table 1 A concept 
 
where ID is the concept unique identifier, label is 
the readable name of the concept,  subclassof indicates 
the relation to another class. As the label has the only 
purpose of highlighting the concept to human readers, 
alternative linguistic expressions are not represented. 
On the contrary, this piece of information is recorded in 
a lexical data base like WordNet. The problem is even 
more obvious when considering relationships.  
 
ID teacherOf 
Label Teaches 
Domain #Faculty 
Range #Course 
 
Table 2 A relationship 
 
In Table 2, domain and range contain the two con-
cepts related to the described binary relation. The label 
teacherOf does not mention alternative linguistic ex-
pressions like: #Faculty gives #Course or #Faculty de-
livers #Course, etc. 
  
For the ontology producers, only one concept or re-
lation name is sufficient. Synonymy is not a relevant 
phenomenon in ontological representations. In fact, it is 
considered a possible generator of unnecessary concept 
name clashes, i.e. concept name ambiguity.  Conceptu-
alisations (as in tables 1,2) are inherently weak when-
ever used to define linguistic models for NLP 
applications. Interpreting questions like: 
 
(1) Who gives/teaches the database class/course 
this year?  
 
with respect to a university domain ontology means 
in fact mapping all the questions onto the concepts and 
relations in Table 2. There is a gap to be filled between 
linguistic and ontological ways of expressing the do-
main knowledge.  
Linguistic interfaces to ontologies 
 In developing an ontological QA system, the main 
problem is to build what we call the “linguistic inter-
face” to the ontology which consists of all the  linguistic 
expressions used to convey concepts and relationships. 
To make this attempt viable, we are currently studying 
methods to automatically relate lexical knowledge bases 
like WordNet (Miller 1995) to domain ontologies 
(Basili et al 2003a) and to induce syntactic-semantic 
patterns for relationships (Basili et al 2003b). 
 
The linguistic interface constitutes the basis on 
which to build the semantic model of the natural lan-
guage processing sub-system. One way of conceiving 
such a model is in terms of syntactic-semantic mapping 
rules that apply to alternative expressions of the same 
conceptual knowledge.  The amount of syntactic analy-
sis such rules  foresee will vary according to the ap-
proach chosen.   
 
Classifying questions  
To facilitate recognition of what are the relevant ex-
pressions to be encoded in the linguistic interface, we 
have introduced a classification of the possible ques-
tions that the system is expected to support. A classifi-
cation often quoted is that in Lauer, Peacocok  and 
Graesser (1992), which mainly builds on speech act 
theory. Another influential, more syntactically-oriented 
approach is that in Moldovan et al. (1999) where to each 
syntactic category correspond one or several possible 
answer types, or focuses (a person, a date, a name, etc.).   
 
Several dimensions have been identified as relevant 
for MOSES 
1. the number of sites and pages in which the an-
swer is to be found. Thus, a first distinction is 
done between site-specific and federated ques-
tions. In the first case, analysis involves only 
one language and one knowledge domain. In 
the second, the interpretation of a question 
produced by a local linguistic analyser is 
matched against the knowledge domain of 
other sites; 
2. sub-domain coverage  (e.g. people, courses, re-
search).  
3. format of the answer: which in MOSES is not 
only a text paragraph as in standard QA, but 
could also be composed of one or more in-
stances of semantic concepts (professors, 
courses) or relations (courses being taught by 
specific professors), whole Web pages, tables, 
etc. due to the heterogeneity of information 
sources  
These dimensions have been explored in “question 
cards” defined by the project’s user groups2.  
 
FORM 1 
Input Hvem underviser i filmhistorie  
(Who teaches film history) 
Syntactic 
type 
Who (Hvem) 
Syntactic 
subtype 
V ≠ copula 
CONTENT  
Focus 
constraint 
Teacher 
Concepts  Faculty 
Course.Name: history of film 
Relations TeacherOf(Faculty, Course) 
Answer 
count  
List 
 
Table 3: Example of question classification 
 
From the point of view of the linguistic analysis, 
however, syntactic category and content are the central 
dimensions of sentence classification. Syntactic catego-
ries are e.g. yes/no question, what-question, who-
question, etc. Subtypes  relate to the position inside the 
question where the focus is expressed, e.g. depending 
on whether the wh-pronoun is a determiner, or the main 
verb is a copula. The content consists of concepts and 
relations from the ontology, the focus constraint3 (the 
ontological type being questioned), and a count feature 
indicating the number of instances to be retrieved. Table 
3 shows an example of linguistic classification. For each 
sentence type, several paraphrases are described.  
 
Ontology Mapping in a Multilingual Envi-
ronment: challenges  
The conceptualisation of the university world  as it 
appears in the DAML+OIL ontology library is an inter-
esting representation for the application scenarios tar-
geted in MOSES (i.e. People/Course/Research).  
Described classes and relations cover in fact, at least at 
a high level, most of the relevant concepts of the  ana-
lysed scenarios.  Such an ontology has been adapted to 
develop conceptualisations for each of the two national 
                                                           
2 The University of Roma III and the Faculty of Hu-
manities at the University of Copenhagen. 
3 In the sense of Rooth (1992). 
university sub-systems (i.e. Italian and Danish) while 
providing additional information required for answering 
the input questions. This is temporal information or 
other kind of information at a border line with the do-
main, (e.g. concepts related to the job market). A first 
important matter we have dealt with is  the language. 
Whereas concept and relation labels in the Italian ontol-
ogy are expressed either in English (for concepts di-
rectly taken from the original source) or in Italian, in the 
Danish counterpart all labels are in Danish. This means 
that a mapping algorithm making use of string similarity 
measures applied to concept labels will have to work 
with translation, either directly between the two lan-
guages involved, or via a pivot language like English. 
The goal would be to establish correspondences such as 
‘Lektor’ a0  (‘AssociateProfessor’) a0  ‘ProfessoreAsso-
ciato’. 
Another problem is related to structural differences: 
not all the nodes in  an ontology are represented also in 
the other and vice-versa,  moreover nodes that are 
somehow equivalent, may have different structural 
placements. This is the case for the ‘Lek-
tor’/’ProfessoreAssociato’ pair just mentioned: in the 
Danish system, ‘Lektor’ is not a subclass of ‘Professor’, 
although “associate professor” is considered a correct 
translation.  
Question analysis  
 
Question analysis is carried out in the MOSES lin-
guistic module associated with each system node. To 
adhere to the semantic Web approach, MOSES poses no 
specific constraints on how the conceptual representa-
tion should be produced, nor on the format of the output 
of each linguistic module. The agent that passes this 
output to the content matcher (an ontology-based search 
engine) maps the linguistic representation onto a com-
mon MOSES interchange formalism (still in an early 
development phase). Two independent modules have 
been developed for Danish and Italian language analy-
sis. They have a similar architecture  (both use preproc-
essing, i.e. POS-tagging and lemmatising, prior to 
syntactic and semantic analyses), but specific parsers. 
Whereas the Danish parser, an adapted version of PET 
(Callmeier 2000) produces typed feature structures 
(Copestake 2002), the Italian one outputs quasi-logical 
forms. Both representation types have proven adequate 
to express the desired conceptual content. As an exam-
ple, the Italian analysis module is described below. 
Analysis of Italian questions 
Analysis of Italian questions is carried out by using 
two different linguistic interpretation levels. The syntac-
tic interpretation is built by a general purpose robust 
syntactic analyser, i.e. Chaos (Basili&Zanzotto 2002). 
This will produce a Question Quasi-Logical Form (Q-
QLF) of an input question based on the extended de-
pendency graph formalism (XDG) introduced in 
(Basili&Zanzotto 2002).  In this formalism, the syntac-
tic model of the sentence is represented via a planar 
graph  where nodes represent constituents and arcs the 
relationships between them. Constituents produced are 
chunks, i.e. kernels of verb phrases (VPK), noun 
phrases (NPK), prepositional phrases (PPK) and adjec-
tival phrases (ADJK). Relations among the constituents 
represent their grammatical functions: logical subjects 
(lsubj), logical objects (lobj), and prepositional modifi-
ers. For example, the Q-QLF of the question 
 
(2) Chi insegna il corso di Database? 
 (Who teaches the database course?) 
 
is shown in Figure 1.  
 
 lsubj lobj di 
NPK NPK VPK PPK 
[Chi] [insegna] [il corso] [di Database][?] 
 Figure 1 A Q-QLF within the XDG formalism 
Then a robust semantic analyser, namely the Dis-
course Interpreter from LaSIE (Humphreys et al. 1996) 
is applied. An internal world model  has been used to 
represent the way in which the relevant concepts (i.e. 
objects) and relationships (i.e. events) are associated 
with linguistic forms (see Figure 2).  Under the object 
node, concepts from the domain concept hierarchy are 
mapped onto synsets (sets of synonyms) in the linguistic 
hierarchy EWN (i.e. the EuroWordNet.base concepts). 
This is to guarantee that linguistic reasoning analysis is 
made using general linguistic knowledge. 
 
Events 
Objects 
Domain 
Concept  
Hierarchy 
WN1.6:EWN 
Base Concepts 
 Figure 2 The world model taxonomy 
 
TEACH_EVENT ==> teach_course. 
teach_course ==> tenere v insegnare v fare. 
 
props(teach_course(E),[ 
 (consequence(E, 
 [relation(E,teacherOf),r_arg1(E,X),r_arg2(E,Z)] ):- 
  nodeprop(E,lsubj(E,X)),  
X <- ewn4123(_),   /* human_1 */ 
  nodeprop(E,lobj(E,Z)),  
Z <- ewn567704(_)  /* education_1 */ 
 ) 
]). 
 
Figure 3 Example of syntactic-semantic inter-
pretation rule 
 
The association of objects and events with linguistic 
forms is used in matching rules as shown in Figure 3. 
The rule expresses the fact that, if any word like tenere, 
insegnare or fare is encountered in relation with a hu-
man_1 (represented by the base concept ewn4123) and 
the word education_1 (ewn567704),  the relation teach-
erOf can be induced.  
 
The analysis resulting for sentence (2) is then: 
 
focus(e2), 
relation(e1,teacherOf), 
r_arg1(e1, person_dch(e2)), 
r_arg2(e1,course_dch(e3)), 
relation(e4,hasSubject), 
r_arg1(e4, course_dch(e3)), 
r_arg2(e4,topic_dch("Database")). 
  
This means that the user is interested in a person, the 
entity e2 of the class person_dch, that is in a relation 
teacherOf with the entity e4 (instance of the class 
course_dch), that is in turn related by hasSubject 
with the topic (i.e. topic_dch) "Database". This result 
can be passed on to the content matcher. 
 
Treating federated questions  
Now we want to extend this approach to question 
analysis in order to manage federated questions. A pos-
sible solution would be sending the natural language 
question to several nodes and let each node interpret it 
against its own domain knowledge. This is unfeasible in 
a multilingual environment. The solution we are inves-
tigating is based on the notion of ontology mapping. Let 
us consider the case of a student questioning   not only 
the Danish but also the Italian site (by selecting specific 
modalities for entering questions):  
 
(3) Hvem er lektor i fransk? 
(Who is associate professor of French?) 
 
As the question is in Danish, it has to be analysed by 
the Danish analysis component, which will produce a 
semantic interpretation roughly corresponding to the 
following term: 
 
all(x) (lektor(x) & CourseOffer(x,y) & 
Course(y) & Name(y, French))4 
 
Since all concepts and relations come from the Dan-
ish ontology, it is not a problem to query the Danish 
knowledge base for all relevant examples. In order to 
query the Italian knowledge base, however, equivalent 
concepts and relations must be substituted for those in 
the “Danish” interpretation. The corresponding Italian 
representation is: 
 
all(x) (ProfessoreAssociato(x) & 
TeacherOf(x,y) & Course(y) &  
Subject(y, French)) 
 
The first problem is establishing a correspondence 
between ‘lektor’ and ‘ProfessoreAssociato’, which as 
shown in the ontology fragments below are not structur-
ally equivalent.  
As suggested in (Pazienza&Vindigni 2003, Med-
che&Staab 2001), equivalence relations must be estab-
lished by considering is-a structures and lexical concept 
labels together. In the example under discussion, an 
initial equivalence can be posited between the top nodes 
of the two ontology fragments, since they both refer 
explicitly to the original DAML+OIL ontology via a 
sameAs relation. However, none of the concept labels 
under ‘Faculty’ in the Italian ontology are acceptable 
translations of ‘Lektor’, nor do any of the nodes refer to 
common nodes in a common reference ontology. Thus, 
the matching algorithm must search further down for en 
equivalent concept by considering possible translations 
of concept labels and testing the relations that equiva-
lence candidates participate in. Thus, distance from a 
common starting node, lexical equivalence and occur-
rence in similar relations are all constraints to be con-
sidered. 
 
                                                           
4 All concepts and relations will in fact be expressed in 
Danish. Here,to facilitate non-Danish readers, we are using 
English equivalents with the exception of the concept ‘Lek-
tor’  under discussion. 
 Lærersta
 
Professorat 
(Professorship) 
Lektor 
(Associate 
Professor) 
Adjunkt 
(Assistant 
 Professor) 
 
… 
Professor 
(FullProfessor) 
GæsteProfessor 
(GuestProfessor) 
 
Faculty 
Professore 
(Tenured 
Professor) 
TitolareCorso 
(Teaching 
Assistant) 
Ricercatore 
(Research 
Assistant) 
 
… 
ProfessoreAssociato 
(Associated 
Professor) 
Ordinario 
(FullProfessor) 
 
 
Figure 4: The “Faculty” Danish and Italian sub-ontologies 
 
The same problem of finding a correct mapping ap-
pears for the relations. In this case, we must be able to 
discover that CourseOffer and TeacherOf  represent the 
same relation. For instance we can rely on the fact that 
they have both two roles, and the concepts filling these 
roles, Faculty and Course (or rather the Danish and Ital-
ian equivalent concepts) correspond. Discovering simi-
larities between relations, however, may be a much 
more complex task than shown in this example. In gen-
eral, it presupposes the ability to map between concepts. 
 
Conclusion  
Our focus in this paper has been, in the context of 
ontology-based QA, to discuss how to interface between 
ontology and linguistic resources on the one hand, and 
ontology and natural language questions on the other 
while remaining within a unique framework. An inter-
esting issue in a multilingual environment is how to 
support questions to federation of sites organised around 
local ontologies.  We have begun to address this issue in 
terms of ontology mapping.  Specific algorithms for 
machine learning and information extraction have also 
been identified and are under development. 
 
References 
Steven Abney (1996) Part-of-speech tagging and 
partial parsing. In G.Bloothooft K.Church, S.Young, 
editor, Corpus-based methods in language and speech. 
Kluwer academic publishers, Dordrecht.  
 
Salah Ait-Mokhtar and Jean-Pierre Chanod. (1997) 
Incremental Finite-state parsing. In Proceedings of 
ANLP97, Washington.  
 
Andreasen, Troels,  Per Anker Jensen, Jørgen F. 
Nilsson, Patrizia Paggio, Bolette Sandford Pedersen and 
Hanne Erdman Thomsen (2002) Ontological Extraction 
of Content for Text Querying, in Natural Language 
Processing and Information Systems, Revised Papers of 
NLDB 2002. Springer-Verlag, pp. 123–136. 
 
Baader, F., D. Calvanese, D. McGuinness, D. Nardi, 
P.F. Patel-Schneinder, eds. (2003) The Description Lo-
gics Handbook: Theory, Implementation, and Applica-
tions, Cambridge University Press 
 
Basili, Roberto, Michele Vindigni, Fabio Massimo 
Zanzotto (2003a) Integrating ontological and linguistic 
knowledge for Conceptual Information Extraction, Web 
Intelligence Conference, Halifax, Canada, September 
2003 
 
Basili, Roberto, Maria Teresa Pazienza, and Fabio 
Massimo Zanzotto (2003b) Exploiting the feature vector 
model for learning linguistic representations of rela-
tional concepts Workshop on Adaptive Text Extraction 
and Mining (ATEM 2003) held in conjuction with Eu-
ropena Conference on Machine Learning (ECML 2003) 
Cavtat (Croatia), September 2003 
 
Basili, Roberto and Fabio Massimo Zanzotto (2002) 
Parsing Engineering and Empirical Robustness Journal 
of Natural Language Engineering 8/2-3 June 2002  
 
Burger, John et al (2002) Issues, tasks and program 
structures to roadmap research in question & answer-
ing (Q&A). NIST DUC Vision and Roadmap Docu-
ments, http://www-nlpir.nist.gov 
/projects/duc/roadmapping.html. 
 
Callmeier, Ulrich (2000) PET – a platform for ex-
perimentation with efficient HPSG processing tech-
niques. In Flickinger, D., Oepen, S., Tsujii, J. and 
Uszkoreit, H. (eds.) Natural Language Engineering. 
Special Issue on Efficient Processing with HPSG. Vol. 
6, Part 1, March 2000, 99–107. 
 
Copestake, Ann (2002) Implementing Typed Feature 
Structure Grammars. CSLI Publications. Stanford Uni-
versity. 
 
Garshol, Lars Marius (2003) Living with Topic 
Maps and RDF. Technical report. 
http://www.ontopia.net/topicmaps/materials/tmrdf.html. 
 
Harabagiu, Sanda,  Dan Moldovan, Marius Paca, 
Rada Mihalcea, Mihai Surdeanu, Rhzvan Bunescu, 
Roxana Girju, Vasile Rus, and Paul Morrescu (2001) 
The role of lexico-semantic feedback in open-domain 
textual question-answering. In Proceedings of the Asso-
ciation for Computational Linguistics, July 2001. 
 
Hobbs, Jerry R., Douglas E. Appelt, John Bear, 
David Israel, Megumi Kameyama, Mark Stickel, and 
Mabry Tyson (1996). FASTUS: A cascaded finite-state 
transducer for extracting information from natural-
language text. In Finite State Devices for Natural Lan-
guage Processing. MIT Press, Cambridge, MA. 
 
Humphreys, K., R. Gaizauskas, S. Azzam, C. 
Huyck, B. Mitchell, H. Cunningham, and Y. Wilks 
(1998) University of sheffield: Description of the 
LASIE-II system as used for MUC-7. In Proceedings of 
the Seventh Message Understanding Conferences 
(MUC-7). Morgan Kaufman, 1998. 
 
Meadche, Alexander and Steffen Staab (2001) Com-
paring Ontologies-Similarity Measures and Comparison 
Study, Internal Report No. 408, Institute AIFB, Univer-
sity of Karlsruhe, Germany, 2001 
 
Miller, George A. (1995) WordNet: A lexical data-
base for English. Communications of the ACM, 
38(11):39--41, 1995. 
 
Pazienza, Maria Teresa and Michele Vindigni 
(2003) Agent-based Ontological Mediation in IE sys-
tems in M.T. Pazienza ed. Information Extraction in the 
Web Era, LNAI 2700, Sprinter Berlin 2003 
 
Rooth, M. (1992) A Theory of Focus Interpretation. 
In Natural Language Semantics, Vol. 1, No. 1, pp. 75-
116. 
 
Voorhees, Ellen M. (2001) The TREC question an-
swering track. Natural Language Engineering 7(4), pp. 
361–378. 
 
Vossen, Piek (1998) EuroWordNet: A Multilingual 
Database with Lexical Semantic Networks Kluwer Aca-
demic Publishers, Dordrecht, October 1998 
 
Woods, W., R. Kaplan, and B. Nash-Weber (1972) 
The Lunar Sciences Natural Language Information Sys-
tem: Final Report. Technical Report, Bolt Beranek and 
Newman, Number 2378, June 1972. 
 
Zajac, Remi (2001) Towards Ontological Question 
Answering, ACL-2001 Workshop on Open-Domain 
Question Answering, Toulose, France, 2001 
