Recently, an overall trend towards increasing complexity of ontologies could be observed, not only in terms of domain modeling, where the complexity should correspond to the information to be modeled, but also as regards the addition of further information, which could be modeled as external resources to the domain model and linked to its relevant elements. This concerns the addition of terminological and linguistic information to the description of classes and properties of ontologies. To respond to this development, we propose a functional approach to the modularization of ontologies, based on terminological, linguistic, and conceptual functions each module ful lls. Only the conceptual elements and their structural properties should remain in the domain model, whereas the formalized terminology and linguistics are described in independent modules referencing the domain models. We provide examples of such complexity in Knowledge Representation systems, discuss related work, and present our approach to modularization in detail.
Nowadays, ontologies in general not only contain domain knowledge but further information central to various tasks of ontology-based systems. For instance, terminological and linguistic details are substantially di erent in nature from the former and usually encoded in labels adjoined to IDs of classes and properties.
There is a growing realization among many researchers that it might not be the best practice to encapsulate such information within the description of classes and properties of domain ontologies. Proposals have already been made for the separation of terminology and lexicon from domain ontologies and for strategies on the linking of this information to the elements of the domain model in a more principled way [1{6]. Our approach to modularization can be considered functional, as it is based on the functions the terminological and linguistic elements used in the context of domain models ful ll. Several tasks such as supporting Information Systems (IS), semantic annotation, lexicographic applications, translation, localization among many others bene t from encapsulated and reusable functions as presented herein.
The need to cull content of labels in ontologies has increased with more possibilities to linguistically process labels, adding linguistic annotations to their textual content and thereby more complexity to the ontology. As a result, reusability and sharing of the information accumulated is considerably impeded since navigation through the entire ontology is required in order to nd linguistically annotated terms that are relevant to ontology-driven applications.
Therefore, following a series of similar proposals [1{3], extending and specifying some points made, we suggest a strict modularization of domain ontologies in a class hierarchy, a terminology, and a linguistic component, all represented in RDF/OWL and related to each other by means of the Simple Knowledge Organization Scheme of the W3C (SKOS) and similar linking mechanisms. Thus, a lexical entry can be used by several terminologies, terms of which are employed in di erent speci c ontologies.
The proposed model largely facilitates the detection of interrelations among ontologies, rendering the formation of new ontologies on the basis of existing independently built ones faster and less complicated, because the model strips ontologies to their core and most essential elements. It equally aims at more compact terminologies and lexicons used in relation with domain modeling, since variants of these can be more easily detected and collapsed onto harmonized sets. Thus, our three-module system represents a mechanism for increasing exibility in reusing ontologies as well as domain-speci c lexicons and terminologies. 2
A class de ned in the RadLex ontology3 serves to exemplify the growing complexity in ontologies. As can be seen in the example below, the class RID 13218 contains all information about its superordinate class and the related properties. Furthermore, information on natural language expressions associated with the class (synonym, NonEnglish Name, Preferred Name, ORIG Preferred Name, Definition) as well as other knowledge sources, i.e., FMAID 67112, were accumulated to form one single ontology class. The knowledge source refers to the Foundational Model of Anatomy (FMA)4. Upon looking at the entry in the FMA ontology, it can quickly be inferred that elements have just been duplicated, such as the de nition, synonym, the (German) Non-English part and the label (preferred name). 3 Version 3, http://bioportal.bioontology.org/ontologies/2027?p=terms 4 The URL for the indicated ID is http://bioportal.bioontology.org/ontologies/ 44507/?p=terms&conceptid=fma\%3AImmaterial_anatomical_entity <name>RID13218</name> <type>anatomy_metaclass</type> <own_slot_value> <slot_reference>FMAID</slot_reference> <value value_type="string">67112</value> </own_slot_value> <own_slot_value> <slot_reference>Synonym</slot_reference> <value value_type="string">immaterial physical anatomical entity</value> </own_slot_value> <own_slot_value> <slot_reference>Non-English_name</slot_reference> <value value_type="string">immaterielles korperliches anatomisches Wesen</value> </own_slot_value> <own_slot_value> <slot_reference>Preferred_name</slot_reference> <value value_type="string">immaterial anatomical entity</value> </own_slot_value> <own_slot_value> <slot_reference>ORIG_Preferred_Name</slot_reference> <value value_type="string">immaterial anatomical entity</value> </own_slot_value> <own_slot_value> <slot_reference>Definition</slot_reference> <value value_type="string">Physical anatomical entity which is a three-dimensional space, surface, line or point associated with a material anatomical entity. Examples: body space, surface of heart, costal margin, apex of right lung, anterior compartment of right arm.</value> </own_slot_value> <own_slot_value>
<slot_reference>Is_A</slot_reference> 4 <value value_type="class">RID13441</value> </own_slot_value> <own_slot_value> <slot_reference>Has_Subtype</slot_reference> <value value_type="class">RID13221</value> <value value_type="class">RID13250</value> <value value_type="class">RID13291</value> <value value_type="class">RID13307</value> <value value_type="class">RID15845</value> <value value_type="class">RID13217</value> </own_slot_value> <own_slot_value> <slot_reference>:ROLE</slot_reference> <value value_type="string">Concrete</value> </own_slot_value> <superclass>RID13441</superclass> </class> [Example of growing complexity in ontologies by means of a RadLex class.]
It seems that the RadLex ontology in this particular case reuses many elements of FMA, as the focus of RadLex is rather on phenomena that can be observed in correlation with speci c organs and not the organs themselves. While this integration of terminological and linguistic knowledge in the eld of anatomy is obviously a good move, re-using established terminology, it appears that it could be more bene cial to provide this pool of information independently from the ontologies modeling the domain. Clear links between the original ontology and terms used as well as linguistic data substantially improve the level of re-usability and readability of semi-structured or de nitional natural language expressions across a large number of ontologies (or taxonomies). 3
Several approaches and models emphasize the importance of separating
conceptual, terminological, and lexical information. Some concentrate on the
terminological aspect [
The Terminae [
A third approach, suggesting the merging of LexInfo and Terminae is CTL
[
Some approaches emphasize the added bene t of a combination of all three
modules for speci c tasks (e.g. [
All approaches above agree that natural language processing and subsequent
linguistic annotation of the terms used in labels are necessary. In order to ensure
interoperability and re-usability, we use standardized models. The
Terminological Markup Framework (TMF), de ned in ISO 16642, ensures the re-usability
of terminological data across applications and the TermBase eXchange (TBX)
format of ISO 30042 represents a best practice for the practical exchange of
terminology. In line with ISO 704, we take a concept oriented approach towards
terminology, de ning terminology as concepts and their designations in a speci c
domain. Consequently, a term is a verbal designation denoting a general concept
in a speci c domain. The lemon model [
We propose LabelNet, a model that modularizes each lexical, linguistic, and terminological function related to ontology labels, establishing a net of interlinked terms with highly detailed information at each level. Term entries in a separate OWL-DL encoded TBX- and TMF-compliant terminology relate semantically to corresponding ontology classes or other conceptual elements and represent the terminological information in detail. Each token5 of every term entry links to a lexical entry, i.e., to a lemma6, syntactic information, and possible additional resources such as further ontologies. Fig. 1 exempli es the structure of LabelNet and shows how each of its modules can be interlinked using SKOS. The example data has been taken from an ontology based on the Belgian National Bank (BNB) taxonomy. Time concepts are linked to the W3C time ontology, e.g., \more than one year" is an interval.
The lexical entries are represented by using partially the lemon model [
By separating the several layers into modules we achieve a more complete and highly detailed perspective of ontology labels. The separation of lexical entries and terms into lexicons and terminologies provides a higher degree of re-usability. In addition, it facilitates a number of computations over these labels, such as the usage of a certain lemma in terms pointing to concepts/role IDs. lemon provides a model that can encode lexical information, using among others RDF, URIs and linking mechanisms, so that language data can be exchanged for example in the Linguistic Linked Open Data cloud7. The model aims at a strict separation of `world knowledge' (describing domain objects that are 5 Tokens can be de ned as all meaningful elements in a text that result from the process of tokenization, i.e., breaking up text into words, phrases, symbols or other meaningful elements. The ordered collection box in Fig. 1 contains lists of tokens as they appear in the terms used in the exempli ed labels. 6 A lemma represents the canonical form of a set of words called lexemes. For example, accrue is the lemma of accrued, accruing, accrues, etc. 7 http://linguistics.okfn.org/resources/llod/ referenced by lexical objects) from `word knowledge' (describing lexical objects). It is itself modular, having a core component that can be supplemented with a set of modules to be used, extended, or ignored as required as illustrated in Fig. 2. For example, a morpho-syntactic module can be attached to the core, specifying speci c values for words used in the term, such as gender (feminine, masculine, neuter), number (singular, plural, dual) and case (nominative, accusative, etc). As this model in essence enables the creation of a lexicon for a given ontology, it is called an ontology lexicon model. lemon as such does not provide an explicit terminological level and refers directly from the lexical entry (in lemon a lexical entry represents the whole content of a label) to an ontology element. In contrast, LabelNet stresses the need and the practicability of a terminological level, we re-use only the non-referential part of the lemon model. 4.2
While lemon o ers a highly interesting perspective, we think that there are still some shortcomings, or possible improvements. A rst case is the fact that lemon supports tokenization of terms included in labels, but not the establishment of the relation between a token represented as a standalone lexical information and the terms in which it can occur. Consequently, we propose an extension that allows for a single lemma to include the information that it is part of a term, in the position speci ed by the tokenization process. Thereby, the word \Verbindlichkeiten" (German for amounts payable or liabilities ), for example, will be linked to a (possibly) substantial number of terms used in various domains (see Fig. 1). In doing so, we can generate a new kind of WordNet, taking into account the inclusion of relevant words in a category of terms. Adopting the idea of lemon, we model only lexical and linguistic information in this separate module, linking to semantic values on the basis of the term ID, which itself links to an ontology element.
As a matter of fact, lemon entries allow only one semantic reference. The lemon model represents the content of labels of one ontology at a time. But frequently one and the same term is used in di erent (even related) ontologies/taxonomies. In this case, two or more lemon entries would be required, leading to redundant lexical/linguistic information only di ering in the entry point to elements of di erent ontologies. One entry pointing to many ontologies represents a more e cient approach. This would also ease generalization over the semantics of such terms.
In case di erent terms are used in concepts of di erent ontologies, but a skos:exactMatch can be established between these concepts, lemon does not provide the means to express the lexical semantic relationship between these terms. As a result, SKOS has to be used as a linking means between those concepts, thereby indirectly establishing the lexical semantic relationship, such as synonymy, between di erent terms.
Apart from linking di erent entries or elements of individual modules, certain constraints need to be re ected. For example, in German and English only the plural of "Verbindlichkeit/liability" might be used within the context of nancial reporting. One possibility in lemon would be to select only terms in which the word "Verbindlichkeit" appears in its plural form "Verbindlichkeiten". Another possibility, which has our preference, would be to associate a feature structure with the lemma we have extracted from the tokens of the ontology labels, in which additional linguistic information can be encoded. Keeping thus the basic lexicon small, i.e., containing mainly lemmas, and using well-de ned feature structures as labels for the edges going from one lemma to a more complex term containing the lemma. We suggest having the constraints expressed in SKOS, linking between a lemma and a term (see Fig. 1): lemma:Verdbindlichkeit -> [plural, feminine, nominative case] -> t1(T3)
The above line expresses that only the plural and nominative form of \Verbindlichkeit", which is feminine, can be used in combination with a term (at least the term \T3"') related to a business reporting ontology. 4.3
Terminologies as such consist of terms denominating concepts, their de nitions
and concept relations. In case of SKOS, these elements are utilized towards
building controlled vocabularies, whereas the TermBase eXchange (TBX) format of
ISO TC 37 can be described as discourse-oriented terminology [
In our model the terminology is supposed to be reusable for other tasks
such as translation, ontology population, ontology building, ontology evolution
to name but a few. Instead of using status attributes such as preferred,
alternative, and hidden, TBX allows for the use of subset information such as project,
application, customer to clarify the di erence between synonyms [
Terminologies provide greater multiplicity than only rdfs:labels. Terms and natural language information acquired for and within the process of ontology building are often lost in the nal representation due to a required univocity of each label. Constructing a net of ontology labels and their synonyms acquired in the building of ontologies and extraction of information results in a domainspeci c, formalized, and reusable resource for ontologies.
Another reason for transferring natural language information from the ontology to terminologies can be found in its ability to represent conceptual relations di erent from ontological relations and thus, enhance the representation of information with linguistic details. For example, a nancial reporting ontology classi es liquid assets as sibling of key balance sheet gures, the latter of which being the parent to assets. In contrast, hypernymic relations in the terminology see assets as top node, whereas liquid assets is one of its children.
TBX is an XML-encoded markup language for the interchange of
terminological information. Due to reasons of cardinality and variation its transformation
to RDF, i.e., SKOS, turns out to be di cult as described in detail in [
Our architectural decisions and selections have been described above, but the speci cation of the process of obtaining each resource and achieving modularization has yet to be detailed. The main input to building the initial ontology is nancial information, such as annual reports of companies, reporting standards (e.g. IFRS, GAAP, XBRL, etc.), stock exchange websites. We extract details from the named sources and build an initial ontology. Furthermore, the extracted information represents the input for the terminology, where all synonyms are depicted. On the basis of the ontology and the terminology, the lexicon is established. So at the core of the following steps lies the formalization of the extracted knowledge in a domain ontology representing our input. 1. Extract labels/terms and linguistic analysis of terms (tokenization, lemmatization, morphological analysis, tagging, parsing, etc). 2. Extract all lemmas, create or map to an existing lemma in a (multilingual) lexicon to collect all lemmas that are used in all possible labels of all possible ontologies. 3. Encode lemmas in lemon. Add a data structure on top of each lemma, which lists all the tokens in all labels in which the lemma is reproduced. This linking also re ects the morpho-syntactic features of the token according to its analysis. 4. Record all morpho-syntactic and lexico-syntactic information and patterns in the corresponding addition to the linguistic module. 5. All identical labels are stored as a unique element in a terminology container.
Specify term entries as to their conceptual relations and establish proper de nitions or adapt de nitions existing in the ontology. 6. Each lemon represented term is associated with a data structure, i.e., terminology, that points to a variety of ontology elements in which those terms have been introduced. 7. Eliminate all the labels and other linguistic information from the ontology, attening class entries to domain speci c details.
As a result, we have two interlinked ontologies of lemmas and terms as used in ontologies/taxonomies. Thereby, we obtain a subset of language data, which is used in domain ontologies. This can be used in order to analyze textual documents and to annotate them semantically, populate ontologies, or support translations with semantics to name but a few. On the other hand, we have a means for testing ontology mapping or merging. 5
The main linking device between ontologies is SKOS, such as the linking between the nancial reporting ontology and the time ontology in the example provided in Fig. 1. Especially with multilingual ontologies the individual concepts and their matching by means of SKOS is important. Oftentimes, the pivotal role of English as a source language leads to translations of labels instead of proper localizations. In case of nancial reporting standards it is indispensable to take local legal and political regulations a ecting the standard into consideration, as the Belgian reporting standard in French might di er substantially from the reporting standard used in France, especially in the use and interpretation of applied French terms.
By conceptualizing the knowledge in each language individually, the ontology is actually created in each language and not simply translated. Thereby, we are in the position of linking for example the English concept pfs_AmountsPayableMore OneYear to the corresponding Italian concept itcc-ci_DebitiEsigibiliOltre EsercizioSuccessivo by employing skos:exactMatch, which implicitly links the term \Debiti Esigibili Oltre l'Esercizio Successivo" to the English term. For existing monolingual ontologies this proposal might serve as a method for merging several monolingual ontologies by establishing links.
The domain ontology represents the starting point for the linking, containing the initial SKOS links to the terminology, as the terminology might be treated as ontology represented in OWL-DL. From the terminology references to the lexicon holding all individual lemmas can be established. At the same time the terminology represents the interface to lexico- and morpho-syntactic patterns as well as syntactical information as such and all tokens, the result from the process of tokenization.
One part of the linking process is the representation of lexico- and morphosyntactic patterns and information to support the evolution and extension of existing domain ontologies. Thereby, the construction of new labels is largely facilitated on the basis of the structure of existing labels.
Syntactic information is represented by combining tokens and dependency
information of individual terms. Basically, syntactic categories are determined
on the basis of part of speech tagging and phrasal categories are used for syntactic
labels. For example N-NP = (length=1, token[
Especially for information extraction in combination with ontology evolution the representation of lexico-syntactic patterns is essential, such as lexicosyntactic ontology design patterns9 and the famous Hearst patterns. One example for their use is the recognition of relations among entities during information extraction. The following sentence has been taken from the International Financial Reporting Standard (IFRS): \The statement of nancial position (sometimes called the balance sheet) includes an entity's assets, liabilities and equity as of the end of the reporting period"10. The lexico-syntactic equivalence <NP class> call in passive <NP class> relation between \statement of nancial position" and \balance sheet" enables us to realize that both terms point to the same ontology concept as synonyms, however, including a description of their di erence in the de nition of the terminology. The Hearst pattern [NP0] [VBG include] [NP1] [NP2]... indicates that \assets, liabilities and equity" can be modeled as subClassOf \statement of nancial position". 6
Modular and encapsulated domain, linguistic, and lexical functions for knowledge modeling enable the support of several IS-related as well as Natural Language Processing (NLP)-driven tasks. Each modularized resource, i.e., ontology, terminology, or lexical information, can either be used as part of the interlinked model we presented or as individual resource for other purposes. One aspect for further improvement certainly is the linking device between the modules, which could be optimized towards an enhanced interoperability with other systems and among the resources themselves.
Acknowledgements. The DFKI part of this work has been supported by the Monnet project (Multilingual ONtologies for NETworked knowledge), co-funded by the European Commission with Grant No. 248458, and by the TrendMiner project, co-funded by the European Commission with Grant No. 287863. 8 http://www.isocat.org/ 9 http://ontologydesignpatterns.org/wiki/Submissions:LexicoSyntacticODPs 10 http://www.ifrs.org/Home.htm