%\documentstyle[epsfig,cupbook,times,mathptm]{book}
%\documentstyle[notitlepage,epsf,cupbook,times,mathptm]{book}
% \input{epsf}
%\renewcommand{\textfraction}{.1}
%\renewcommand{\topfraction}{.9}
% \sloppy

%%\hyphenation{DPDAs}

\documentstyle[cupbook, epsfig]{book} 
\begin{document}
\cupbookchapter
{Vossen}{Condensed Meaning In EuroWordNet}
{Piek Vossen}{Condensed Meaning In EuroWordNet}
%\author%
%{
% P. Vossen
% % University of Amsterdam \\
% % The Netherlands \\
% % {\tt piek.vossen@hum.uva.nl}
%}

\begin {abstract}
This paper discusses condensed meaning in the EuroWordNet project. 
In this project several wordnets for different languages are combined in a multilingual database.
The matching of the meanings across the wordnets makes it necessary to account for polysemy in a generative way 
and to establish a notion of equivalence at a more global level. Finally, we will describe an attempt to set up a 
more fundamental ontology which is linked to the meanings in the wordnets as derived complex types. 
The multilingual design of the EuroWordNet database makes it possible to specify how the lexicon of 
each language uniquely maps onto these condensed types.
\end {abstract}

\section{Introduction}
\label{intro}

The aim of EuroWordNet\footnote {EuroWordNet is funded by the EC as projects LE2-4003 and LE4-8328. It is a joint enterprise of the University of Amsterdam (co-ordinator), the Fundacion Universidad Empresa (Madrid and Barcelona), Istituto di Linguistica Computazionale del CNR (Pisa), University of Sheffield, University of Tuebingen, University of Avignon, University of Tartu, University of Brno, Bertin (Paris), Memodata (Avignon), Xerox Research Center (Grenoble) and Lernout and Hauspie (Antwerp). Further information on the project can be found at: http://www.hum.uva.nl/\~ewn.} is to develop a multilingual database with wordnets in several European languages: English, Dutch, Italian, Spanish, French, German, Czech, and Estonian. Each language-specific wordnet is structured along the same lines as WordNet \cite {Miller90}: synonyms are grouped into synsets, which are related by means of basic semantic relations such as hyponymy (e.g. between {\it car} and {\it vehicle}) or meronymy relations (e.g. between {\it car} and {\it wheel}). By means of these relations all words are interconnected, constituting a huge network or wordnet. Since the lexicalization of concepts is different across languages, each wordnet in the EuroWordNet database represents an autonomous and unique system of language-internal relations. This means that each wordnet represents the lexical semantic relations between the lexicalized words and expressions of the language only: no artificial classifications (such as External-Body-Part, Inanimate\-Object) are introduced to impose some structuring of the hierarchy \cite{Vossen98,VossenBloksma98}. In addition to the relations between the synsets, each language-synset is related to an Inter-Lingual-Index (ILI), connecting all these wordnets. In the database, it is possible to go from synsets in one language to synsets in any other wordnet that are linked to the same ILI-records and to compare the lexical semantic structures. This is illustrated in Figure 1 for the language-specific synsets linked to the ILI-record {\it drive}. The ILI is an unstructured fund of concepts, mainly based on the synsets taken from WordNet1.5 and adapted to provide a better matching across the wordnets. Each ILI-record consists of a synset, a gloss specifying the meaning and a reference to its source. No relations are maintained between the ILI-records as such. The development of a complete language-neutral ontology is considered to be too complex and time-consuming given the limitations of the project. As an unstructured list there is no need to discuss changes or updates to the index from a many-to-many perspective and it is easier to deal with complex mappings of meanings across wordnets. Furthermore, it will be possible to indirectly see a structuring of a set of ILI-records by viewing the language-internal relations of the language-specific concepts that are related to the set of ILI-records. It is thus possible to get any hierarchical structuring, according to any ontology or wordnet that is linked to the index (including WordNet1.5). In \cite {Vossen97,PetersVossenDiezAdriaens98}, further details are given how lexical semantic structures or configurations can be compared in the EuroWordNet database. 

Some language-independent structuring of the ILI is still provided by two separate ontologies, which may be linked to ILI records (see Figure 1):

\begin {itemize}
\item the Top Concept ontology, which is a hierarchy of language-independent concepts, reflecting important semantic distinctions, e.g. {\it Object} and {\it Substance}, {\it Dynamic} and {\it Static}
\item a hierarchy of domain labels, which are knowledge structures grouping meanings in terms of topics or scripts, e.g. {\it Traffic}, {\it Road-Traffic}, {\it Air-Traffic}, {\it Sports}, {\it Medical} 
\end {itemize}

Both the Top Concepts and the domain labels can be transferred via the equivalence relations of the ILI-records to the language-specific meanings, as illustrated in Figure 1. The Top Concepts {\it Dynamic} and {\it Location} are for example directly linked to the ILI-record {\it drive} and therefore indirectly also apply to all language-specific concepts related to this ILI-record. Via the language-internal relations the Top Concepts can be further inherited to all other related language-specific concepts. The main purpose of the ontologies is to provide a common framework for the most important concepts in all the wordnets.

\begin{figure}[t]
\setlength{\unitlength}{0.000790in}%
\begin{center}
\begin{picture}(3533,1806)(2768,-3973)
\thicklines
\large
\end{picture}
\end{center}
\caption[]{The global architecture of the EuroWordNet database}
\label{architecture}
\end{figure}

The development of the individual wordnets takes place at different sites, using separate tools and databases. To achieve maximal compatibility of the independently-created results, we have determined a set of common Base Concepts. These Base Concepts play a major role in the separate wordnets, which is measured in terms of the number of relations with other concepts and their position in the hierarchy.\footnote {Base Concepts should not be confused with the Basic Level concepts defined by \cite {Rosch77}. Base Concepts are typically more general and abstract, which is a natural consequence of the criteria for their selection. So whereas {\it table} may a Basic Level Concept we expect {\it furniture} to be a Base Concept } Each site has made a selection in the local wordnets, and the translations of these selections into equivalent WordNet1.5 synsets have been merged to form a set of 1024 fundamental concepts (both nominal and verbal), which is shared by all sites. The further development of the wordnets takes place by encoding the language-internal relations for this set and extending it top-down to more specific meanings that depend on it. A special effort has been made to manually encode the relations for these concepts in a coherent and compatible way, so that the cores of the different wordnets are of sufficient quality. For that purpose, we have classified these Base Concepts in terms of the Top-Ontology mentioned above \cite {Rodriguez98}.

The wordnets are (as much as possible) built from existing resources that still have a traditional structuring of senses against which the generative approach opposes \cite {Pustejovsky95}. This means that different contextual interpretations are sometimes enumerated as distinct senses and in other cases collapsed or only partially represented. We see for example that WordNet1.5 only lists one meaning for {\it embassy} as the {\it building}, whereas the Dutch wordnet gives separate meanings for the {\it building} and the {\it institute} for the equivalent {\it ambassade}. In the case of {\it school} both wordnets list both senses, but the Italian wordnet only represents the {\it institute}. These inconsistencies make clear that the comparison of wordnets in the database will confront us with many situations in which meaning interpretations are distributed in different ways. To be able to make a decision over the compatibility of such mismatches it is necessary to define a common policy for the treatment of sense-enumeration in the individual wordnets. This will be discussed in section 2 that deals with complex types in the individual wordnets. By specifying disjunctive and conjunctive combinations of interpretations or conceptualizations it is possible to represent particular types of polysemy in a more condensed way. Nevertheless, other types of regular polysemy will still occur and in section 3 we will discuss how these can be grouped by introducing complex types in the Inter-Lingual-Index itself. Different partial reflections of interpretation across wordnets can then still be matched via more coarse meanings in the ILI that connect several of these interpretations. Finally, the top-ontology concepts, mentioned above, can be combined to form complex combinations that can be applied to ILI-records, and thus indirectly provide a more systematic potential for interpretations for the synsets in the local wordnets linked to these ILI-records.

\section {Complex types in the language-specific wordnets}
\label {section2}
In WordNet1.5 multiple hyperonyms incidentally occur. The next example {\it spoon} has two hyperonyms {\it cutlery} and {\it container} that apply to the same synset:

\begin{description}
\item [spoon] -- (a piece of cutlery with a shallow bowl-shaped container and a handle; used to stir or serve or take up food):
\item [-{\sc has\_hyperonym}: cutlery]-- (implements for cutting and eating food)
\item [-{\sc has\_hyperonym}: container]-- (something that holds things, especially for transport or storage)
\end{description}

In EuroWordNet we systematically use this option to encode combinations of meaning aspects in a single sense, resulting in complex types. Following \cite {Vossen95}, three different complex types are distinguished, depending on the combinatorial constraints of the hyperonyms:

\begin {itemize}
\item disjunctive hyperonyms
\item conjunctive hyperonyms
\item non-exclusive hyperonyms
\end {itemize}

Disjunctive hyperonyms are incompatible types that never apply simultaneously. We find many examples of these disjunctions among so-called {\it functionals}, which are nouns that refer to the participant in an event but do not restrict for the type of entity participating. The typical definition pattern for these nouns is a disjunction of genus words as shown in the following examples taken from the Longman Dictionary of Contemporary English \cite {Procter78}:

\begin {description}
\item[arrival]-- a person or thing that arrives or has arrived
\item[puzzler]-- a person or thing that puzzles
\item[threat]-- a person, thing or idea regarded as a possible danger
\end {description}

The disjunction represents an open range of entity types that can play a certain role in the event. Alternatively, dictionaries may also enumerate different senses for the disjuncted types as is illustrated by the different senses of {\it threat} in WordNet1.5:

\begin {description}
\item [menace, threat1] -- (something that is a source of danger)
\item [-{\sc has\_hyperonym}: danger] -- (a cause of pain or injury or loss; he feared the dangers of travelling by air;)[-{\sc has\_hyperonym}: causal agent, cause, causal agency] -- (any entity that causes events to happen)
\item [terror, scourge, threat4] -- (a person who inspires fear or dread)
\item [-{\sc has\_hyperonym}: person, individual, someone, mortal, human, soul] -- (a human being)
\end {description}

Here we see that sense 1 of {\it threat} is linked to a non-restrictive class {\it danger} which can be any {\it causal agent} (see also the head of the gloss: {\it something}), whereas sense 4 expresses the same concept restricted to {\it person}. By distinguishing different synsets, it is suggested that these are also distinct concepts or senses, which is highly doubtful. As the more generic sense already suggests, anything can be a {\it threat} which means that we could need any number of synsets to express this range. Cases such as {\it threat} will in principle be encoded in EuroWordNet as a single sense using a so-called role-relation, possibly extended with a disjunctive range of hyperonyms:

\begin {description}
\item [threat]
\item [-{\sc role\_agent}] threaten
\item [-{\sc has\_hyperonym}] person; {\em disjunctive}
\item [-{\sc has\_hyperonym}] thing; {\em disjunctive}
\item [-{\sc has\_hyperonym}] idea; {\em disjunctive}
\end {description}

The status of the multiple hyperonyms is indicated by the label {\em disjunctive} that can be added to a relation. Not all cases of disjunctive hyperonyms in definitions also represent an open range of types, as is shown in the following examples from LDOCE:

\begin {description}
\item [acquittal 2]-- the act of declaring or condition of being found not guilty
\item [adherence 2]-- the act or condition of sticking to something firmly
\end {description}

In these examples a fixed choice is given as the interpretation of the derived nominal between {\it act} and {\it condition}. The combination of both is incompatible: the former denotes a dynamic event and the latter a static situation. Since we are not dealing with an open range of interpretations we have to assume that they represent different senses. Because the polysemy is regular it is still possible to generalize over this relation. This is however captured outside the language-specific wordnet, as will be explained in the following section.

Disjunctive ranges are not restricted to hyponymy relations only. In the following examples from WordNet1.5 we see that different senses of {\it door} are distinguished depending on the type of whole of which it is a part:

\begin {description}
\item [door 1] -- (a swinging or sliding barrier that will close the entrance to a room or building; he knocked on the door; he slammed the door as he left)
\item [-{\sc has\_holonym}: doorway, door, entree, entry, portal, room access] -- (the space in a wall through which you enter or leave a room or building; the space that a door can close; he stuck his head in the doorway)
\item [doorway, door 2 , entree, entry, portal, room access] -- (the space in a wall through which you enter or leave a room or building; the space that a door can close; he stuck his head in the doorway)
\item [-{\sc has\_holonym}: wall] -- (a partition with a height and length greater than its thickness; used to divide or enclose or support)
\item [door 6]-- (a swinging or sliding barrier that will close off access into a car; she forgot to lock the doors of her car)
\item [-{\sc has\_holonym}: car, auto, automobile, machine, motorcar] -- (4-wheeled; usually propelled by an internal combustion engine; he needed a car to get to work)
\end {description}

The different holonyms of which {\it door} can be a PART are listed here as different synsets (sense 1 and 6 of {\it door}), suggesting that these reflect different concepts. Again, there could be any number of holonyms that have a {\it door} as a part. According to the principle that is applied here these would all result in different synsets for {\it door}. In EuroWordNet, these meanings are combined in a single complex type listing the disjunctive holonyms in which it can be incorporated:

\begin {description}
\item[door]
\item [-{\sc has\_holonym}]car; {\em disjunctive}
\item [-{\sc has\_holonym}]airplane; {\em disjunctive}
\item [-{\sc has\_holonym}]room; {\em disjunctive}
\item [-{\sc has\_holonym}]building; {\em disjunctive}
\end {description}

Note that in the opposite situation conjunction of meronyms is the default:

\begin {description}
\item [car]
\item [-{\sc has\_meronym}]door; {\em conjunctive} \footnote {In general, absence of a label triggers the default interpretation in EuroWordNet. It is therefore not necessary to explicitly indicate {\em conjunction} of {\it meronyms} in the database as it is done in this example.}
\item [-{\sc has\_meronym}]wheel; {\em conjunctive}
\end {description}

Conjunctive hyperonyms always apply simultaneously and thus can never be incompatible. Typical examples of conjunctive hyperonyms are found for specific lexicalizations of verbs in which multiple aspects are combined. In Dutch, we see for example that many verb compounds combine a resultative verb and a manner of motion and thus can be classified by both hyperonyms:

\begin {description}
\item [doodschoppen] to kick to death
\item [-{\sc has\_hyperonym}] doden (to kill); {\em conjunctive}
\item [-{\sc has\_hyperonym}] schoppen (to kick); {\em conjunctive}
\item [opentrekken] to pull open
\item [-{\sc has\_hyperonym}] openen (to open); {\em conjunctive}
\item [-{\sc has\_hyperonym}] trekken (to pull); {\em conjunctive}
\end {description}

Here the conjunctive label indicates that both aspects are always implied in the meaning of the verb. In many other cases, the hyperonyms are non-exclusive: both aspects may apply simultaneously or one of both may apply:

\begin {description}
\item [knife]
\item [-{\sc has\_hyperonym}]weapon
\item [-{\sc has\_hyperonym}]cutlery
\end {description}

Absence of the disjunctive/conjunctive label expresses optionality and compatibility of the perspectives. 

Finally, disjunction and non-exclusiveness also apply in the reversed direction to the hyponyms of a class. As argued by both \cite {VossenBloksma98} and \cite {Guarino98}, there is an important difference between co-hyponyms that represent disjunct classes, such as {\it cat} and {\it dog} and hyponyms that can cross-classify with these disjunct types, such as {\it pet} and {\it draught animal}. As a {\it dog} it is impossible to have {\it cat}-specific properties, but it is not unlikely that it has {\it pet} or {\it draught animal} properties. A similar observation can be made with respect to the substitution behavior of these co-hyponyms. We can refer to {\it dogs} with {\it pet} and {\it draugh animal} but certainly not with {\it cat}. In this respect, the non-disjunct classes resemble the {\it functionals} discussed above, being orthogonal to the disjunct types. Since all relations in EuroWordNet are encoded separately in both ways, it is possible to use the same label {\em disjunctive} to differentiate the hyponyms if a class in the opposite direction (see below).

Nevertheless, there may be strong stereotypical preferences or restrictions to some types. The class of {\it pets} is in our culture typically associated with {\it cats} and {\it dogs} but not with {\it insects}. On the one hand we thus may want to express that the class of {\it pets} has {\it cats} and {\it dogs} as members but we do not want to classify {\it cat} and {\it dog} by all possible orthogonal classes that exist. There is a difference in status between the conventional classification of {\it dog} and {\it cat} as an {\it animal} and the classification as a {\it pet}. The latter classification is more relevant to {\it pet} than for the members. To deal with this phenomenon, a separate label {\em reversed} is used to mark the implicational direction of a relation. The next examples then shows the result when we combine the different kind of labeling and directions of relations that have been discussed:

\begin {itemize}
\item animate being
\subitem {\sc has\_hyponym} plant; {\em disjunctive}
\subitem {\sc has\_hyponym} person; {\em disjunctive}
\subitem {\sc has\_hyponym} animal; {\em disjunctive}
\subitem {\sc has\_hyponym} parent
\subitem {\sc has\_hyponym} winner
\subitem {\sc has\_hyponym} favorite
\item animal
\subitem {\sc has\_hyponym} horse; {\em disjunctive}
\subitem {\sc has\_hyponym} cat; {\em disjunctive}
\subitem {\sc has\_hyponym} dog; {\em disjunctive}
\subitem {\sc has\_hyponym} pet
\subitem {\sc has\_hyponym} draught animal
\item pet
\subitem {\sc has\_hyperonym} animal
\subitem {\sc has\_hyponym} cat
\subitem {\sc has\_hyponym} dog
\item draught animal
\subitem {\sc has\_hyperonym} animal
\subitem {\sc has\_hyponym} horse
\subitem {\sc has\_hyponym} draught dog
\item cat
\subitem {\sc has\_hyperonym} animal
\subitem {\sc has\_hyperonym} pet; {\em reversed}
\item dog
\subitem {\sc has\_hyperonym} animal
\subitem {\sc has\_hyperonym} pet: {\em reversed}
\subitem {\sc has\_hyperonym} draught animal; {\em reversed}
\item draught dog
\subitem {\sc has\_hyperonym} dog; {\em conjunctive}
\subitem {\sc has\_hyperonym} draught animal; {\em conjunctive}
\item horse
\subitem {\sc has\_hyperonym} animal
\subitem {\sc has\_hyperonym} draught animal; {\em reversed}
\subitem {\sc has\_hyperonym} riding animal; {\em reversed}
\end {itemize}

Hyponyms that are not labeled as {\em disjunct} are thus orthogonal to the other co-hyponyms. Cross-classification is only excluded by explicit marking as {\em disjunct} classes. The label {\em reversed} here indicates that the conceptual implication is the other way around. A {\it horse} is a typical example of a {\it draught animal} but {\it draught animal} is not a typical class of {\it horse}. In other words: the link from {\it draught animal} to {\it horse} is needed to define or explain the usage of the former and not the latter. Finally, we see that, in the case of {\it draught dog}, the link to {\it draught animal} is indeed necessary to define it and consequently we have two hyperonyms that are {\em conjunctively} combined. This illustrates the difference in status of the multiple hyperonyms for {\it horse} and for {\it draught dog}.

By means of disjunctive/conjunctive/reversed hyp(er)onyms it is thus possible to encode several generative aspects of meaning: open ranges of entities (disjunctive types) and different types of parallel classifications (multiple hyperonyms). In many other cases, the wordnets will however still contain polysemous entries, listing incompatible hyponymic relations. These will be discussed in the next section.

\section {Complex types in the Inter-Lingual-Index}
\label {section3}
We mentioned before that incompatible hyperonyms that do not represent an open range of entities are kept separate in the wordnets. Among these are many cases of regular polysemy that apply to larger groups of conceptual classes, e.g. the {\it building/institute} polysemy mentioned in the introduction. As suggested, inconsistent listing of the regular polysemy in the wordnets may result in a situation that the local synsets of, for example {\it university}, cannot be matched across wordnets. \cite {HampFeldweg97} describe how such polysemy can be accounted for in a specific wordnet for German. To limit this danger in EuroWordNet for all involved wordnets, we extend the ILI with globalized senses that represent sets of more specific but related senses of the same word. In Figure 2, we see that the original linking of Dutch, Italian and Spanish equivalents for {\it university} have been extended with an {\em eq\_metonym} relation to a new so-called {\em Composite} ILI-record {\it university}, which contains a reference to two more specific meanings. Via the {\em eq\_metonym} relations the synsets can be retrieved despite of the different ways in which they are linked to the more specific synsets. It is not necessary that the metonymy-relation also holds in the local language. In this example only the Dutch wordnet has two senses that parallel the metonymy-relation in the ILI. The Italian and Spanish example only list one sense (which may be correct or an omission in their resources). In the case of Spanish there are multiple equivalences to both senses of {\it university}, whereas the Italian synset is only linked to the {\it building} sense. The Spanish synset is in fact equivalent to the new Composite ILI-record.

\begin{figure}[t]
\setlength{\unitlength}{0.000790in}%
\begin{center}
\begin{picture}(3533,1806)(2768,-3973)
\thicklines
\large
\end{picture}
\end{center}
\caption[]{Inter-Lingual-Index with a Composite ILI-record for {\it university}\ }
\label{composite}
\end{figure}

As a side effect, the relation between the two Dutch senses is now expressed via the metonymy-equivalence relation to the more global Composite ILI-record. The Composite ILI-record may also create metonymic relations between different forms that represent the same semantic relation, such as ''universiteit" (university institute) and ''universiteitsgebouw" (university building) in Dutch. This shows that productive interpretations from a complex type such as {\it university} can have different realizations in languages. Whereas one language may express these interpretations with the same form (resulting in polysemy) another language may use different compounds or derivations.

Similar globalized records are added for {\em generalizations}. Generalizations apply to explicit enumeration of meaning specialization. For example, in the Dutch resource there is only one sense for ''schoonmaken" (to clean) which simultaneously matches with at least 4 senses of {\it clean} in WordNet1.5:

\begin {itemize}
\item make clean by removing dirt, filth, or unwanted substances from
\item remove unwanted substances from, such as feathers or pits, as of chickens or fruit
\item remove in making clean; Clean the spots off the rug
\item remove unwanted substances from - (as in chemistry)
\end {itemize}

The senses in WordNet1.5 do not seem to be incompatible but merely differentiate different contexts and objects that are involved (without being complete). The Dutch equivalent could be used in any of these situations as well. Such extreme proliferation of meaning occurs quite often in WordNet1.5. On the one hand this is a good feature for an interlingua, because we can very precisely establish equivalence relations, on the other hand, it makes it more difficult to comprehensively establish equivalence between synsets across the wordnets. It is clear that arbitrary listings of such specializations may lead to similar situations as described above that synsets across wordnets are linked to different specializations and thus cannot be matched. By introducing a single generalized meaning for cases such as {\it clean} that groups these specializations, and by adding an {\sc eq\_generalization} relations from the relevant synsets in the wordnets to the generalization we will achieve a better matching. 

Finally, there is a specific globalization relation, {\sc eq\_diathesis}, to capture sense differentiation due to diathesis alternations \cite {Levin93}, e.g.:

\begin {description}
\item [Italian: cambiare (to change)] 1 intransitive (to become different)
2 transitive (to make different)
\item [Dutch: bewegen (to move)] 1 intransitive (to change place or position)
2 transitive (to cause to change place or position)
6 reflexive ((of people, animals) to change place or position)
\end {description}

Here we see that Italian ''cambiare" 1 and 2 (change) exhibit a transitive/intransitive alternation which correlates with a difference in causation. Something similar holds for different senses of ''bewegen" (move) in Dutch, which refer as intransitives verbs to a non-causative change-of-position and as transitives to the causation of such a change (this also holds for ''mover" (move) in Spanish and ''muovere" (move) in Italian). This phenomenon is very wide-spread in all the resources that are used, including WordNet1.5 which forms the basis for the ILI. By adding a single meaning that abstracts from causation (and thus from transitive or intransitive realization) it is possible to get a more systematic encoding on the one hand and provide a more stable matching across the wordnets as well.

The generation of these equivalence relations is to a large extent done automatically (see \cite {PetersVossenDiezAdriaens98,PetersPetersVossen98} and \cite {Buitelaar98}). After extending the ILI with more global concepts, the {\sc eq\_metonym}, {\sc eq\_generalization} or {\sc eq\_diathesis} will be automatically generated for all synsets that have at least one of the specific ILI-records in the globalized ILI-records as the target of an {\sc eq\_synonym} or {\sc eq\_near\_synonym} relation. There is no need for the local wordnet builders to consider each of these equivalence-extensions manually. Note that this procedure will also generate a relation when interpretations are represented by different word forms in languages, as has been suggested for Dutch compounds such as {\it universiteitsgebouw} (university building). This brings in a new perspective in the discussion. If realization of generative meanings is not limited to polysemy, it is less clear where one should stop. In theory, there could be a vague language that uses a few forms and extreme degree of polysemy, or a language which uses an extreme amount of forms, each with an unique interpretation. In the interlingua we could build in any level of sense-grouping or globalization but it is not clear what is a natural or practical level per se. The granulation of the ILI is now determined by the English language (as represented by WordNet1.5), which has a tendency to use polysemy rather than derivation or compounding (as in e.g. more Germanic languages).

\section {Complex types of Top-Concepts}
\label {section4}

The above discussion showed that a well-designed interlingua or language-neutral ontology may have many benefits from which all the linked wordnets can profit. We also demonstrated that the separation of the interlingua from the language-specific realizations may help to clarify the way meaning is proliferated in the lexicalized vocabulary of languages. Still, the sense-groupings in the ILI mainly serve a practical purpose: to improve the matching of synsets across the wordnets. As we explained in the introduction, it is not feasible to develop a complete language-neutral ontology within the limits of this project, and without knowing in what respects languages differ. The ILI is therefore still nothing but an index for matching meanings, without much structure. This being said, we nevertheless have began to provide a more fundamental specification of the meanings in the form of a top-ontology of 63 semantic distinctions. This ontology has been applied to a set of 1024 so-called Base Concepts (BCs) that play a fundamental role in establishing the relations in the different wordnets \cite {Rodriguez98}. These Base Concepts are represented as ILI-records and thus indirectly give access to the linked synsets in the local wordnets.

The first starting point for the top ontology is that the wordnets are linguistic ontologies, representing the lexicalization patterns of languages. We therefore used semantic distinctions which are common in linguistic paradigms: Aktionsart models \cite {Vendler67,Verkuyl72,Dowty79,Verkuyl89,Pustejovsky91}, entity-orders \cite {Lyons77}, Aristotle's Qualia-structure \cite {Pustejovsky95}. The second starting point is that the ontology should reflect the diversity of the set of common BCs. In this sense, the classification of the common BCs in terms of the top-concepts should result in homogeneous Base Concept Clusters with an average size. Large clusters will be further subdivided and very small clusters will be generalized. Finally, we can mention as important characteristics:

\begin {itemize}
\item the Top Concepts are hierarchically ordered by means of a subsumption relation but there can only be one super-type linked to each Top Concept: multiple inheritance between top-concepts is not allowed.
\item in addition to the subsumption relation, Top Concepts can have an opposition-relation to indicate that certain distinctions are disjoint, whereas others may overlap.
\item there may be multiple relations from ILI-records to Top Concepts. This means that the BCs can be cross-classified in terms of multiple Top Concepts (as long as these are not disjoint): i.e. multiple inheritance from Top Concept to Base Concept is allowed.
\end {itemize}

It is important to realize that the Top Concepts (TCs) are more like semantic features than common conceptual classes. We typically find TCs for {\em Living} and for {\em Part} as meaning components or facets but we do not find a TC {\em Bodypart}, even though this may be more appealing to a non-expert. BCs representing {\it body parts} are now cross-classified by the conjunction of {\em Living} and {\em Part}. The reason for this is that the diversity of the BCs would require many cross-classifying concepts where {\em Living} and {\em Part} are combined with many other TCs. Furthermore, it turned out that the BCs typically abstract from particular features but these abstractions do not show any redundancy: i.e. it is not the case that all things that are {\em Living} also always share other features. An explanation for the diversity of the BCs is the way in which they have been selected. To be useful as a classifier or category for many concepts (one of the major criteria for selection) a concept must capture a particular generalization but abstract from (many) other properties. Likewise we find many classifying meanings which express only one or two TC-features but no others. In this respect the BCs typically abstract one or two levels from the cognitive Basic-Level as defined by \cite {Rosch77}.

Following \cite {Lyons77} we distinguish at the first level 3 types of entities:

\begin {description}
\item [1stOrderEntity (always concrete nouns)] Any concrete entity (publicly) perceivable by the senses and located at any point in time, in a three-dimensional space.
\item [2ndOrderEntity (nouns, verbs and adjectives)] Any Static Situation (property, relation) or Dynamic Situation, which cannot be grasped, heard, seen, felt as an independent physical thing. They can be located in time and occur or take place rather than exist; e.g. continue, occur, apply
\item [3rdOrderEntity (always abstract nouns)] An unobservable proposition which exists independently of time and space. They can be true or false rather than real. They can be asserted or denied, remembered or forgotten. E.g. idea, thought, information, theory, plan. 
\end {description}

Since the number of 3rdOrderEntities among the BCs was limited compared to the 1stOrder and 2ndOrder Entities we have not further subdivided them. The 1stOrderEntities and 2ndOrderEntities are further subdivided according to the hierarchy given in \ref {topontology}, where the superscripts indicate the number of assigned BCs. For a more complete description of the BCs and TCs see \cite {Rodriguez98}. Here will only discus the most important distinctions.

\subsection { Classification of 1st-Order-Entities}
\label {section41}

The 1stOrderEntities are distinguished in terms of four main ways of conceptualizing or classifying a concrete entity:

\begin {description}
\item [Origin] the way in which an entity has come about.
\item [Form] as an a-morf substance or as an object with a fixed shape, hence the subdivisions Substance and Object.
\item [Composition] as a group of self-contained wholes or as a part of such a whole, hence the subdivisions Part and Group.
\item [Function] the typical activity or action that is associated with an entity.
\end {description}

These classes are comparable with Aristotle's Qualia roles as described in Pustejovsky's Generative lexicon, (the {\em Agentive role}, {\em Formal role}, {\em Constitutional role} and {\em Telic Role} respectively: \cite {Pustejovsky95}) but are also based on our empirical findings to classify the BCs. BCs can be classified in terms of any combination of these four roles. 

\begin{figure}[t]
\setlength{\unitlength}{0.000790in}%
\begin{center}
\begin{picture}(3533,1806)(2768,-3973)
\thicklines
\large
\end{picture}
\end{center}
\caption[]{The EuroWordNet Top-Ontology}
\label{topontology}
\end{figure}

The main-classes are then further subdivided, where the subdivisions for {\em Form} and {\em Composition} are obvious given the above definition, except that {\em Substance} itself is further subdivided into {\em Solid}, {\em Liquid} and {\em Gas}. In the case of {\em Function} the subdivisions are based only on the frequency of BCs having such a function or role. In principle the number of roles is infinite but the above roles appear to occur more frequently in the set of common Base Concepts. Finally, a more fine-grained subdivision has been made for {\em Origin}, first into {\em Natural} and {\em Artifact}. The category {\em Natural} covers both inanimate objects and substances, such as stones, sand, water, and all living things, among which animals, plants and humans. The latter are stored at a deeper level below {\em Living}. The intermediate level {\em Living} is necessary to create a separate cluster for natural objects and substances, which consist of {\it living material} (e.g. {\it skin}, {\it cell}) but are not considered as {\it animate beings}. Non-living and natural objects and substances, such as natural products like {\it milk}, {\it seeds}, {\it fruit}, are classified directly below {\em Natural}.As suggested, each BC that is a 1stOrderEntity is classified in terms of these main classes. However, whereas the main-classes are intended for cross-classifications, most of the subdivisions are disjoint classes: a concept cannot be an {\em Object} and a {\em Substance}, or both {\em Natural} and {\em Artifact}. This means that within a main-class only one subdivision can be assigned. Consequently, each BC that is a 1stOrderEntity has at least one up to four classifications:

\begin {description}
\item [fruit] Comestible (Function); Object (Form); Part (Composition); Plant (Natural, Origin)
\item [skin] Covering (Covering); Solid (Form); Part (Composition); Living (Natural, Origin)
\item [cell] Part (Composition); Living (Natural, Origin)
\item [life 1] Group (Composition); Living (Natural, Origin)
\item [reproductive structure 1 ]Living (Natural, Origin)
\end {description}

Finally, with respect to {\em Composition} it needs to be said that only concepts that essentially depend on some other concept, are classified as either {\em Part} or {\em Group}. It is not the case that all persons will be classified as Parts because they may be part of group. {\em Group}, on the other hand, typically depends on the elements as part of its meaning. The default interpretations is therefore to be an independent {\em Whole}.

\subsection { The classification of 2ndOrderEntities}
\label {section42}

As explained above, 2ndOrderEntities can be referred to by nouns and verbs (and also adjectives or adverbs) denoting static or dynamic Situations, such as {\it birth}, {\it live}, {\it life}, {\it love}, {\it die} and {\it death}. All 2ndOrderEntities are classified using two different classification schemes, which represent the first division below 2ndOrderEntity:

\begin {description}
\item [SituationType] the event-structure in terms of which a situation can be characterized as a conceptual unit over time
\item [SituationComponent] the most salient semantic component(s) that characterize(s) a situation
\end {description}

{\em SituationType} reflects the way in which a situation can be quantified and distributed over time, and the dynamicity that is involved. It thus represents a basic classification in terms of the event-structure (in the formal tradition) and the predicate-inherent Aktionsart properties of nouns and verbs. The {\em SituationComponents} represent a conceptual classification, resulting in intuitively coherent clusters of word meanings. The {\em SituationComponents} reflect the most salient semantic components that apply to our selection of Base Concepts. Examples of {\em SituationComponents} are: {\em Location}, {\em Existence}, {\em Cause}.

Typically, {\em SituationType} represents disjoint features that cannot be combined, whereas it is possible to assign any range or combination of {\em SituationComponents} to a word meaning. Each 2ndOrder meaning can thus be classified in terms of an obligatory but unique {\em SituationType} and any number of {\em SituationComponents}. Following a traditional Aktionsart classification \cite {Vendler67,Verkuyl72,Dowty79,Verkuyl89}, {\em SituationType} is first subdivided into {\em Static} and {\em Dynamic}, depending on the dynamicity of the Situation:

\begin {description}
\item [Dynamic] Situations implying either a specific transition from one state to another (Bounded in time) or a continuous transition perceived as an ongoing temporally unbounded process; e.g. event, act, action, become, happen, take place, process, habit, change, activity. Opposed to {\em Static}.
\item [Static] Situations (properties, relations and states) in which there is no transition from one eventuality or situation to another: non-dynamic; e.g. state, property, be. Opposed to {\em Dynamic}.
\end {description}

In general words, {\em Static Situations} do not involve any change, {\em Dynamic Situations} involve some specific change or a continuous changing. {\em Static Situations} are further subdivided into {\em Properties}, such as length, size, which apply to single concrete entities or abstract situations, and {\em Relations}, such as distance, space, which only exist relative to and in between several entities (of the same order):

\begin {description}
\item [Property] Static Situation which applies to a single concrete entity or abstract Situation; e.g. colour, speed, age, length, size, shape, weight.
\item [Relation] Static Situation which applies to a pair of concrete entities or abstract Situations, and which cannot exist by itself without either one of the involved entities; e.g. relation, kinship, distance, space.
\end {description}

{\em Dynamic Situations} are subdivided into events which express a specific transition and are bounded in time ({{\em BoundedEvent}), and processes which are unbounded in time ({\em UnboundedEvent}) and do not imply a specific transition from one situation to another (although there can be many intermediate transitions):

\begin {description}
\item [BoundedEvent] Dynamic Situations in which a specific transition from one Situation to another is implied; Bounded in time and directed to a result; e.g. to do, to cause to change, to make, to create.
\item [UnboundedEvent] Dynamic Situations occurring during a period of time and composed of a sequence of (micro-)changes of state, which are not perceived as relevant for characterizing the Situation as a whole; e.g. grow, change, move around, live, breath, activity, hobby, sport, education, work, performance, fight, love, caring, management.
\end {description}

The {\em SituationComponents} divide the Base Concepts into conceptually coherent clusters. The set of distinctions is therefore based on the diversity of the set of common Base Concepts that has been defined. As far as the set of Base Concepts is representative for the total wordnets, this set of {\em SituationComponents} is also representative for the whole. As said above, a verb or 2ndOrder noun may thus be composed of any combination of these components. However, it is obvious that some combinations make more sense than others. The more specific a word is, the more components it incorporates. Just as with the 1stOrderEntities we therefore typically see that the more frequent classifying nouns and verbs only incorporate a few of these components. In the set of common Base-Concept, such classifying words are more frequent, and words with many {\em SituationComponents} are therefore rare. Below are some examples of typical combinations of {\em SituationComponents}:

\begin {itemize}
\item Experience + Stimulating + Dynamic+Condition (undifferentiated for Mental or Physical)
\subitem Verbs: cause to feel unwell; cause pain
\item Physical + Experience + SituationType (undifferentiated for Static/Dynamic)
\subitem Nouns: sense; sensation; perception; 
\subitem Verbs: look; feel; experience;
\item Mental + (BoundedEvent) Dynamic + Agentive
\subitem Verbs: identify; form an opinion of; form a resolution about; decide; choose; understand; call back; ascertain; bump into; affirm; admit defeat
\subitem Nouns: choice, selection
\item Mental + Dynamic + Agentive
\subitem Verbs: interpret; differentiate; devise; determine; cerebrate; analyze; arrange
\subitem Nouns: higher cognitive process; cerebration; categorization; basic cognitive process; argumentation; abstract thought
\item Mental + Experience + SituationType (undifferentiated for Static/Dynamic)
\subitem Verbs: consider; desire; believe; experience
\subitem Nouns: pleasance; motivation; humor; feeling; faith; emotion; disturbance; disposition; desire; attitude
\item Relation+Physical+Location 
\subitem Verbs: go; be; stay in one place; adjoin 
\subitem Nouns: path;course; aim; blank space; degree; direction; spatial relation; elbow room; course; direction; distance; spacing; spatial property; space
\end {itemize}

The 1stOrder en 2ndOrder TCs can thus be combined in a partial lattice. Combinations of TCs have been applied to the BCs to get the most specific description that was still considered to be valid in the local wordnets. This resulted in 124 clusters for the 1stOrderEntities (491 BCs) and 314 clusters for the 2ndOrderEntities (500 BCs). The TC-combinations can be seen as complex types assigned to the ILI-records that represent the BCs. These types do not represent metonymic relations as discussed in the previous section but more direct presuppositions or entailments. They predict a range of interpretations corresponding to the combination of TCs. Possibly, more complex types can be derived for more specific concepts than the BCs but this has not been realized in EuroWordNet. In principle, the complex types should carry over to all the synsets in the local wordnets that have a (direct) equivalence relation with the BC. The implications can then be further distributed to all the hyponyms of the concepts in the local wordnets. As we have discussed, multiple hyperonyms can result in complex types as well, and thus also in further combinations of these basic implications in the local wordnet. Figure 4 illustrates this for {\it spoon} in WordNet1.5, which inherits via the multiple hyperonyms both the single feature {\em Function} from {\it container} and the feature-combination {\em Artifact} and {\em Object} from {\it artifact object}. In this example there is some redundancy because of the diamond hyponymy-structure.

However, it is important to realize that the Top Ontology does not necessarily correspond with the language-internal hierarchies. Each language-internal structure has a different mapping with the top-ontology via the ILI-records to which they are linked as equivalents. For example there are no words in Dutch that correspond with technical notions such as 1stOrderEntity, 2ndOrderEntity, 3rdOrderEntity,\footnote {The opposite situation also occurs. Words such as ''iets" (a thing, event or idea; anything) or ''zaak" (thing or concept) are more abstract than the top-level concepts.} but also not with more down-to-earth concepts such as the functional 1stOrder concept {\em Container}. These levels will thus not be present in the Dutch wordnet. From the Dutch hierarchy it will hence not be possible to simply extract all the {\it containers} because no Dutch word meaning is used to group or classify them. Nevertheless, the Dutch {\it containers} may still be found either via the equivalence relations with English {\it containers} which are stored below the sense of {\it container} or via the TopConcept clustering {\em Container} that is imposed on the Dutch hierarchy (or any other ontology that may be linked to the ILI). Figure 4 shows a fragment of the Dutch wordnet in which there is no equivalent for {\it container} nor for {\it artifact object} and {\it artifact}. 

\begin{figure}[t]
\setlength{\unitlength}{0.000790in}%
\begin{center}
\begin{picture}(3533,1806)(2768,-3973)
\thicklines
\large
\end{picture}
\end{center}
\caption[]{Complex types in the Top Ontology and the local wordnets}
\label{containers}
\end{figure}

The dotted line from the ILI-record {\it container} to ''lepel" (spoon), indicates a so-called {\sc eq\_has\_hyperonym} relation to a more general ILI-record that can additionally classify the concept. This information can either be transferred from WordNet1.5 or it can explicitly be encoded to represent the Base Concept {\it container} in Dutch in the form of all the important hyponyms. This demonstrates another possibility to combine top-ontology types in the local wordnet. If a language-specific concept is more complex, inclusive than any of the ILI-records it is possible to link it to multiple ILI-records via the {\sc eq\_has\_hyperonym} relation. Any top-concept associated to these ILI-records will transfer to the language-specific meaning.

This discussion clearly shows that the realizations of interpretations, even from a generative point of view, is a matter of the language. Even though, we demonstrated that is possible to set up a powerful and predictive system for deriving complex meanings, still, each language represents a unique lexical mapping to these meanings or aspects of these meanings.

\section {Conclusion}
\label {conclusion}

It will be clear that both the individual wordnets and the multilingual database as a whole will profit from a generative approach, which reduces the (often inconsistent) enumeration of interpretations and improves the mapping across languages. For the separate wordnets, the possibility of disjunctive and conjunctive hyperonyms or hyponyms (and holonyms or meronyms) makes it possible to capture the different facets of meaning in a single sense or synset. For the multilingual database as a whole, the more coarse sense groups in the ILI will provide a smoother matching of meanings across the wordnets. We have seen that different lexicalizations of meaning aspects, either via polysemy, compounding or derivation, can thus still be interconnected. Finally, we described the EuroWordNet top-ontology, which represents a more fundamental language-independent structuring of the ILI in the form of complex types. Although this top-ontology has stronger predictive power, it is nevertheless limited to a smaller set of so-called Base Concepts that play a major role in the individual wordnets. These implications can however still be carried over to the specific meanings in the local wordnets via the system of language-internal relations.

The current top-ontology is just a basic lattice of distinctions, in which the axioms and implications are not further formalized. The current ontology does not formally express how for example {\em Function} relates to {\em 2ndOrderEntities} and we thus cannot differentiate Action-Result patterns of polysemy or derivation from Action-Agent patterns. The flexibility of the lattice was needed during the devlopment phase and the formalization was not necessary for encoding the basic semantic relations in the wordnets. However, we would like to further extend this lattice with a formalization in the near future. This would more fundamentally explain how lexicalizations of meanings in the vocabularies of languages are related to more fundamental complex types, and consequently, give a better context to predict other extensions or derivations of meanings. 

\begin{thebibliography}{}
\makeatletter
%\show\@lbibitem
\makeatother

\bibitem[\protect\citename{Bloksma {\it et al.}\ } 1996] {BloksmaDiezVossen96} Bloksma, L., Diez-Orzas, P., Vossen, P. (1996). {\it User requirements and functional specification of the EuroWordNet project}, EuroWordNet deliverable D001, Amsterdam, University of Amsterdam.

\bibitem[\protect\citename{Buitelaar} 1998] {Buitelaar98} Buitelaar, P. (1998). {\it Corelex: Systematic Polysemy and Underspecification}, PhD, Brandeis University.

\bibitem[\protect\citename{Dowty} 1979] {Dowty79} Dowty, D.R., (1979). {\it Word meaning and Montague grammar: the Semantics of Verbs and Times in Generative Semantics and in Montague's PTQ}, Dordrecht: Reidel.

\bibitem[\protect\citename{Hamp and Feldweg} 1997] {HampFeldweg97} Hamp, B., Feldweg, H. (1997). GermaNet: a Lexical-Semantic Net for German. In Vossen, Calzolari, Adriaens, Sanfilippo, Wilks (eds), {\it Proceedings of the ACL/EACL-97 workshop on Automatic Information Extraction and Building of Lexical Semantic Resources for NLP Applications}, Madrid.

\bibitem[\protect\citename{Guarino} 1998] {Guarino98} Guarino, N. (1998). Some Ontological Principles for Designing Upper Level Lexical Resources. In A. Rubio, N. Gallardo, R. Catro and A. Tejada (eds) {\it Proceedings of First International Conference on Language Resources and Evaluation}, Granada, 527-534.

\bibitem[\protect\citename{Levin} 1993] {Levin93} Levin, B. (1993). {\it English Verb Classes and Alternations}, Chicago, University of Chicago Press.

\bibitem[\protect\citename{Lyons} 1977] {Lyons77} Lyons, J. (1977). {\it Semantics}, London, Cambridge University Press.

\bibitem[\protect\citename{Miller {\it et al.}\ } 1990] {Miller90} Miller, G., Beckwith, R., Fellbaum, C., Gross, D., and Miller, K. (1990). {\it Five Papers on WordNet}, Princeton, CSL Report 43. Cognitive Science Laboratory, Princeton University.

\bibitem[\protect\citename{Peters, Peters and Vossen} 1998] {PetersPetersVossen98} Peters, W., Peters, I., and Vossen, P. (1998). The Reduction of Semantic Ambiguity in Linguistic Resources. In A. Rubio, N. Gallardo, R. Catro and A. Tejada (eds) {\it Proceedings of First International Conference on Language Resources and Evaluation}, Granada, 409-416.

\bibitem[\protect\citename{Peters {\it et al.}\ } 1998] {PetersVossenDiezAdriaens98} Peters, W., Vossen, P., Diez-Orzas, P., Adriaens, G. (1998). Cross-linguistic Alignment of Wordnets with an Inter-Lingual-Index. In: Nancy Ide, Daniel Greenstein, Piek Vossen (eds), {\it Special Issue on EuroWordNet. Computers and the Humanities} 32, 221-251.

\bibitem[\protect\citename{Procter} 1978] {Procter78} Procter, P. (ed.) (1978). {\it Longman Dictionary of Contemporary English}, Harlow and London, Longman.

\bibitem[\protect\citename{Pustejovsky} 1991] {Pustejovsky91} Pustejovsky, J. (1991). The syntax of event structure. {\it Cognition} 41, 47-81.

\bibitem[\protect\citename{Pustejovsky} 1995] {Pustejovsky95} Pustejovsky, J. (1995). {\it The Generative Lexicon}, Cambridge, MA: MIT Press.

\bibitem[\protect\citename{Rodriguez {\it et al.}\ } 1998] {Rodriguez98} Rodriguez, H., Climent, S., Vossen, P., Bloksma, L., Roventini, A., Bertagna, F., Alonge, A., Peters, W. (1998). The Top-Down Strategy for Building EuroWordNet: Vocabulary Coverage, Base Concepts and Top Ontology. In Nancy Ide, Daniel Greenstein, Piek Vossen (eds), {\it Special Issue on EuroWordNet. Computers and the Humanities} 32, 117-152.

\bibitem[\protect\citename{Rosch} 1977] {Rosch77} Rosch, E. (1977). Human Categorisation. In N. Warren (eds) {\it Studies in Cross-Cultural Psychology} I, London, Academic Press, 1-49.

\bibitem[\protect\citename{Vendler} 1967] {Vendler67}Vendler, Z. (1967). {\it Linguistics and philosophy}, Ithaca, Cornell University Press.

\bibitem[\protect\citename{Verkuyl} 1972] {Verkuyl72} Verkuyl, H. (1972). {\it On the compositional nature of the aspects}, Dordrecht,  Reidel.

\bibitem[\protect\citename{Verkuyl} 1989] {Verkuyl89} Verkuyl , H. (1989). Aspectual classes and aspectual distinctions. {\it Linguistics and Philosiphy} 12, 39-94.

\bibitem[\protect\citename{Vossen} 1995] {Vossen95} Vossen, P., (1995). Grammatical and Conceptual Individuation in the Lexicon, PhD. {\it Studies in Language and Language Use} 15, Amsterdam, IFOTT, University of Amsterdam. 

\bibitem[\protect\citename{Vossen {\it et al.}\ } 1997] {Vossen97} Vossen, P., Diez-Orzas, P., Peters, W. (1997). The Multilingual Design of EuroWordNet. In: P. Vossen, N. Calzolari, G. Adriaens, A. Sanfilippo, Y. Wilks (eds) {\it Proceedings of the ACL/EACL-97 workshop Automatic Information Extraction and Building of Lexical Semantic Resources for NLP Applications}, Madrid, 1-8.

\bibitem[\protect\citename{Vossen and Bloksma} 1998] {VossenBloksma98} Vossen P., and Bloksma, L., (1998). Categories and classifications in EuroWordNet. 
In A. Rubio, N. Gallardo, R. Catro and A. Tejada (eds) {\it Proceedings of First International Conference on Language Resources and Evaluation}, Granada, 399-408.

\bibitem[\protect\citename{Vossen} 1998] {Vossen98} Vossen, P. (1998). Introduction to EuroWordNet. In Nancy Ide, Daniel Greenstein, Piek Vossen (eds), {\it Special Issue on EuroWordNet. Computers and the Humanities} 32, 73-89.

\end{thebibliography}{}
%\bibliographystyle{namelabel}

%%\endinput
\end{document}