Saying is not Modelling

Christophe Roche

Condillac-Listic, University of Savoie, Campus Scientifique

73376 Le Bourget du Lac cedex, France

Abstract: In this article we claim that the conceptual modelling built from text

is rarely an ontology. Such a conceptualization is corpus-dependent and does

not offer the main properties we expect from ontology, e.g. reusability and

soundness. Furthermore, ontology extracted from text in general does not match

ontology defined by expert using a formal language. Such a result is not

surprising since ontology is an extra-linguistic conceptualization whereas

knowledge extracted from text is the concern of textual linguistics.

Incompleteness of text and using rhetorical figures, like synecdoche, deeply

modify the perception of the conceptualization we may have. It means that

ontological knowledge, which is necessary for text understanding, is not in

general embedded into documents. The article will end on some remarks about

formal languages. If they allow to define “a specification of a

conceptualization” they nevertheless raise their own issues mainly due to their

epistemological neutrality. Ontology design remains an epistemological issue.

1 Introduction

Whatever their domain: information systems, databases, natural language processing,

knowledge based systems, etc. applications are more and more ontology-oriented

[1],[2],[3]. Such a success is mainly due to what ontology

promises; it means a way

of capturing and representing a shared understanding of a domain that can be

understood and used by humans as well as by software. Then, one of the main

problems to be solved is to build domain ontology.

Ontology building, as knowledge base building, is a difficult and time-consuming

task. It is the reason why a lot of works are currently done on ontology acquisition

(e.g. Ontology learning ECAI workshops, KCAP workshops). Since we can consider

that technical and scientific documents convey some domain knowledge, ontology

building can rely on knowledge acquisition from texts [4].

Ontology building from text corresponds to a “knowledge reverse engineering from

text” process described by the following figure.

There is today an agreement on the definition of ontology. We can resume the most of

definitions by saying that: “an ontology is a shared description of concepts and relationships

of a domain expressed in a computer readable language”.

Roche C. (2007).

Saying is not Modelling.

In Proceedings of the 4th International Workshop on Natural Language Processing and Cognitive Science, pages 47-56

DOI: 10.5220/0002426000470056

 SciTePress

Fig. 1. Ontology building from text.

This process raises several questions. The main one is “Do ontology built from text

and ontology defined by experts in a formal language match?”

In other words what

are the consequences for ontology building of using a given language, either natural

or formal? What is information lost – and what is information introduced – when

experts write text in natural language? On the other hand, do – and how – formal

languages influence conceptualization? This article will try to give some answers to

these questions.

2 Ontology Reverse Engineering from Text

2.1 Industrial Context

In order to illustrate our talk, let us take an industrial application carried out by

Ontologos corp.

for EDF Research & Development (Electricity of France) [5], [6].

The main goal of this application was the re-appropriation of the ontology describing

the concepts in the field of control and instrumentation. The problem is all the more

difficult in that this knowledge is not directly accessible in usable form but is spread

out through various bodies of knowledge, and more especially into documents. In

particular we worked on a corpus

about relay. The ontology of relay we have defined

is currently used for different applications including a content management system for

document classification and information retrieval.

The difference between these ontologies appears through the expressions “linguistic

ontology” and “formal ontology”.

http://www.ontologos-corp.com

A corpus is a collection of texts which have been selected according to some criteria [7] and

for a given objective. In the framework of our application, the criteria were mainly:

“produced by a same community of practice”, “about a same topic”, “belonging to the same

type of text (descriptive)”, “under the same form”.

2.2 Lexicon and Lexical Structure

The goal of the first stage is to build a lexical structure. It means to build a network of

words linked by linguistic relationships; where words – in general nouns or noun

phrases – denote concepts

and linguistic relationships are mainly hyponymy,

synonymy and meronymy relationships. Thus the first step is to extract candidate

terms for concept’s names. Extracting candidate terms and linguistic relationships

from corpus by automatic text analysis is today an active research domain [4], [8], [9].

Statistical methods based on Harris’s distributional hypothesis, i.e. collocation

analysis of terms [10], as well as linguistic methods based for example on regular

expressions can be used. Regular expressions like “adjective noun” and “noun noun”

patterns allow the extraction of expressions like “electromagnetic relay”, “threshold

relay”, “on/off relay”, “voltage relay”, “undervoltage relay”, ‘overvoltage relay”, etc.

from the relay corpus. The result, which must be validated by experts, is a lexicon of

words of usage considered as many as possible concept’s names.

This lexicon is structured according to linguistic relationships like hypernymy (versus

hyponymy), synonymy, meronymy and so on. Here too, these linguistic relationships

can be ‘automatically’ extracted from the corpus using both a syntactic analysis – “a

voltage relay is a kind of relay” – and the lexical structure of noun phrases. For

example, linguistic expressions made up of several words with the same ending (i.e.

ending with the same words, for example with the same noun) give interesting

information about the structure of the lexicon. The following linguistic expressions

“voltage relay”, “threshold relay”, “electromagnetic relay” can be considered as many

as hyponyms of “relay” (let us recall that in this article, linguistic expressions are

given between quotation marks).

Fig. 2. A part of the lexical structure.

2.3 Conceptual Structure and Ontology

The second stage is to deduce the conceptual structure from the lexical one. If we

assert that a term

denotes a concept and the hyponymy relationship is a linguistic

In this article, names (nouns and noun phrases) as well as linguistic expressions will be

written between quotation marks: “relay”, “voltage relay”, “A voltage relay is a kind of

relay.” etc.; while concepts will be written between the lower and upper symbols: the name

“relay” denotes the concept <relay>.

A term can be considered as a lexicalised concept like in Wordnet [11] where a concept is

defined as a synset, i.e. a set of synonymous words.

translation of the subsumption relationship, then the noun phrase “voltage relay”

denotes the concept <voltage relay> which is a sub concept of the concept <relay>.

The result is a conceptual structural which matches the lexical one.

Fig. 3. The conceptual structure defined from the lexical structure.

The last, but not least, stage is the validation of the conceptual structure. During this

stage, the conceptual structure is completed as necessary, concept’s names are

normalized and words of usage are associated with concepts. The result is the

ontology of domain. In our example, since writing technical documents relies on the

terminology

of the domain, the previous conceptual structure has been labeled as a

valid ontology of the domain.

3 Does It Really Work?

The process of ontology building from text is quite clear and well defined, and a lot of

very interesting work is currently being done on these different stages. But is it so

simple (even if statistical and linguistic methods can be very complex)? And does

such ontology really satisfy our expectations?

3.1 Application-Oriented Validation

One generally says that an ontology is defined for a given goal [1]. Let us take the

example of an ontology-oriented content management system. As a lot of ontologies

built from text, the ontology of relay is both used for classifying documents

and

information retrieval [12]. Thus, search for information about a concept, for example

<threshold relay>, must return all information about this kind of relay and about its

more specialised concepts (using if necessary the ‘is a’ inheritance relationship for

inferences). But in our case, no information about <voltage relay> is returned when it

should do so. For experts, all information about <voltage relay> concerns <threshold

relay>. Although the conceptualization is not wrong since it has been validated by the

experts (a <voltage relay> is really a kind of <relay>), it is not completely correct.

Where is the problem?

A terminology can be viewed as a constraint language based on a normalized vocabulary.

A document is classified on every concept the content of document refers to.

3.2 Ontology from Text and Ontology from Expert

Ontology built from text (also called “linguistic ontology”) is corpus-depend, which is

not surprising. It means that even when different communities of practice share a

same reality (for example between users and suppliers about relay), it is not possible

to define a sharable and reusable ontology from text as far as these communities use

their own language (Language for Special Purpose). It is the reason why some say

that “these ontologies are domain- and task-specific: concept definitions result from

the selection of a single interpretation context that reflects the application

requirements; they are intended to reflect one of the ways knowledge can be

perceived through the use of language in documents” [8].

In order to better understand the problem we were faced with, experts were asked

to define directly their domain conceptualization in a formal language, independently

of the words of usage

. The result is a formal ontology quite different from text-

oriented ontology. An ontology-dedicated language based on the specific-difference

theory was used by the experts. In such a theory a concept is defined from a

previously defined concept by indicating its specific difference. The concept’s names

are arbitrary and can be normalized. Thus, for experts the concept <voltage relay> is

not a kind of <relay> at the same level as the concept <on-off relay> or the <threshold

relay> one. It is a kind of <threshold relay> whose threshold value is voltage. The

final ontology, described below, does not match the ontology built from text. But it

can be shared and reused between the different communities of practice.

Fig. 4. A part of the formal ontology.

4 Saying is not Modelling

Ontology acquisition from text relies on a set of strong hypotheses. The first of these

is to say that experts can translate ontological knowledge within a corpus which more

or less defines a closed world containing all the necessary information. The second

hypothesis states that the reverse process is possible, based on the fact that a noun, or

a noun phrase, denotes a concept and the hyponymy relationship is a linguistic

translation of the subsumption one. The last hypothesis is to say that the conceptual

structure matches the lexical structure and the former can be deduced from the latter.

Validation by experts allows to bring up the conceptual structure as a domain

ontology. But ontology built from text in general does not satisfy our expectations in

terms of sharing, reusability, consensus and soundness. The problems we encounter

By ‘words of usage’ we mean words which are used for writing texts.

are mainly due to the fact that these hypotheses are too strong and not always true,

even if we use constrained languages for writing technical document. In reality,

ontology building from text is the concern of textual linguistics. One of the latter’s

principles is the incompleteness of text. This implies that understanding text, and then

understanding the meaning of terms, requires extra-linguistic knowledge which by

definition is not included in the corpus. The ‘closed world’ hypothesis is not stricto

sensu applicable. Furthermore text is written for a given goal according to a given

writer’s intention. All this information is not included in corpus: “the meaning (in

general) of a sign needs to be explained in terms of what users of the sign do (or

should) mean” [13].

Let us go back to the lexical structure (fig. 2) and the formal ontology (fig. 4). The

noun phrase “voltage relay” really denotes a concept in the formal ontology. In fact it

is a shortcut, a word of usage of the terminology, for the more complete expression

“threshold relay whose threshold value is voltage”. “Voltage relay” denotes the

concept <voltage threshold relay> of the formal ontology whose name can be

normalized as “voltage threshold relay”. This example illustrates the fact that using

rhetorical figures, like metonymy

or ellipse, and in particular synecdoche

is a very

ordinary practice in writing technical documents. Such rhetorical figures assume that

both writers and readers share a same and pre-existent conceptualization of the

world which is necessary for understanding meaning of term and text. This pre-

existent knowledge – which is the domain ontology – is not included in texts and then

can not be extracted from them. Only some relationships are explicit; and if these

relationships can always be justified in their context, they can be undesirable from the

domain conceptualization point of view. This is the case for the conceptual structure

built from the lexical one (fig. 3) where a <voltage relay> concept has been defined

and linked by a ‘is a’ relationship with the <relay> concept when no relationship has

been extracted with the <threshold relay> concept. These rhetorical figures refer to

implicit knowledge, either concepts or relationships, which is not possible to know,

except for experts. Let us precise that experts have validated the ontology built from

text (fig. 3) for the same reasons they use synecdoche figures, in reference to the

implicit knowledge of their domain. Natural language is not a suitable language for

specifying conceptualization: it is not its aim; as well as the main goal of writing text

is not to define ontology. Knowledge acquisition from text corresponds to a

semasiologic approach where we first find terms (mainly nouns) and then define them

in a given context. Such an approach is concerned with linguistics and more precisely

with lexical semantics [14] whose main objective is word meaning

. It is then

difficult to reuse and share such a contextual knowledge. Nevertheless, a lot of useful

information can be extracted from text; especially if one considers than an ontology is

also a vocabulary of terms

with their definition: “An [explicit] ontology may take a

A metonymy is a figure of speech in which one word is substituted for another with which it

is closely associated.

A synecdoche is a figure of speech in which a part is used for the whole.

Linguistics is mainly interested in the relationships between signifier and signified when

ontology is mainly interested in the relationships between concept and object.

It is important to bear in mind that words of usage of LSP (Language for Special Purpose)

and terms of terminology are not necessary the same.

variety of forms, but necessarily it will include a vocabulary of terms and some

specification of their meaning (i.e. definitions).” [15]. The main result of knowledge

acquisition from text is a network of words of usage linked by linguistic relationships.

There is no concept in text, and the lexical structure does not match with the domain

conceptualization.

5 Formal Languages

In theory, domain ontology represents task-independent, and then sharable and

reusable, knowledge of a domain. A formal approach should allow to reach such a

goal. As a matter of fact, natural language, even constrained in its syntax and

semantics, cannot be used for concept definitions. We need a formal language for the

definition of a conceptualization – such a “specification of a conceptualisation” is

called an ontology [16] –. This is a useful means to avoid the issues raised by natural

language and to reach agreement: if you accept the hypothetical and deductive

approach of the formal system, you are obliged to accept its constructions, i.e. the

domain conceptualization.

Nevertheless, the famous hypothesis of Sapir and Whorf [17], [18], concerning the

interdependence of thought and language, is also applicable to formal languages. This

means that the choice of the formal language for the definition of concepts is

important. The way an ontology is built and the way a concept is defined directly

depends on the formal language which is used; and the results will not be the same.

Today formal languages are mainly logic-oriented. The concepts are represented as

unary predicates when their attributes, or slots, are represented as binary predicates,

also called roles. Description logic [19] is a good example of logic appropriated to

knowledge representation. On the other hand, frame representation languages [20], in

spite of the criticism of [21], are semi-formal and more human-readable languages.

They allow to define concepts as a set of slots and organize them according to an

inheritance and hierarchical relationship. OWL, the web ontology language [22],

combines the advantages of these two approaches. It is a dedicated language for

building ontology based on the W3C philosophy and description logic while

providing a human readable formalism with the Protégé environment. The final

ontology depends on the formal language which is used. So, the frame-oriented

ontology of our relay example might be very similar to the conceptual model

extracted from the lexical structure if we stay too close to text. On the other hand, the

logic-oriented ontology of the relay, if it is different in its expression, does not solve

our problem if we forget to explicitly express the relationship between <voltage

relay> and <threshold relay>. In fact, knowledge representation requires a formal but

also an epistemological oriented language. It means a language which can help the

knowledge engineer to capture the nature of knowledge: for example a set is not a

concept, even if a concept can be interpreted as the set of its subsumed objects. As a

matter of fact, logic is a neutral (or flat) language which cannot represent the different

kinds of knowledge: a unary predicate can represent either a concept or a property

while binary predicates can represent either attributes (internal relationships) or

relationships between concepts (external relationships). Some interesting work has

been done in order to introduce epistemological principles in logic, for example the

‘ontological rigidity’ constraint [23], [24]. But such principles do not really define

guidelines for ontology building in the sense that they do not help the knowledge

engineer to identify and structure concepts, they only constrained value of well

formed formula.

6 The OK Model

The OK (for Ontological Knowledge) language is an ontology-oriented language

which relies on epistemological and formal principles [25]. It is based on the specific-

difference theory. This theory considers a conceptualization as a system of concepts

organized according to their differences more than factoring attributes shared by

objects: a concept is defined from a previously existing one by adding a specific

difference. The difference is then the main principle of the ontology building process

on which identifying and structuring concepts rely on. Let us also remark that the

agreement problem is reduced to the agreement on differences.

Fig. 4. The OCW

environment for ontologies by specific differentiation.

Coming back to our example of ‘relay’, this approach reveals that the difference

between the concept denoted by the linguistic expression “voltage relay” and the one

denoted by “threshold relay” is the threshold value. There is no a <voltage relay>

concept (fig. 3) but a <voltage threshold relay> concept (fig. 6) whose linguistic

expression “voltage relay” is one of its possible words of usage. Furthermore, in

technical domains a lot of experts agree with the classical (Aristotelian) definition of a

concept: a <voltage threshold relay> is a <threshold relay> whose threshold value is

voltage, and then on the resulting conceptual structure.

“Ontology Craft Workbench”. Ontology written in OK can be translated in OWL [26].

7 Conclusion

Since we can consider that technical and scientific documents convey some domain

knowledge, ontology building can rely on knowledge acquisition from texts. But such

a conceptualization is corpus-dependent and does not offer the main properties we

expect from ontology, e.g. reusability and soundness. Furthermore, ontology extracted

from text in general does not match ontology defined by expert using a formal

language.

The knowledge extracted from text is a linguistic knowledge. The lexical structure

is a network of words of usage linked by linguistic relationships like hyponymy,

synonymy, etc. This lexical structure is a linguistic “picture” of the domain ontology.

A picture built in a given and particular context, for a given goal which reflects a

particular linguistic use of the domain conceptualization. As a matter of fact, texts fall

within language in action. Using rhetorical figures, such as metonymy and ellipse, is a

very ordinary practice even for writing technical documents. Such figures of speech

assume that both writers and readers share a same and pre-existent conceptualization

of the world which is necessary for understanding meaning of words of usage and

texts. This pre-existent knowledge – which is the domain ontology – is not included in

texts and then can not be extracted from them. A conceptual model “directly” built

from a lexical structure will probably not be reusable, because too corpus-dependent,

and then probably not correct if we consider an ontology as a non-contingent

knowledge. The lexical structure and the domain ontology do not match.

At last, if formal languages allow to define “a specification of a conceptualization”

they nevertheless raise their own issues mainly due to their epistemological neutrality.

Ontology design remains an epistemological issue which requires epistemological-

oriented languages.

References

1. Staab, S., Studer, R.: Handbook on Ontologies. Steffen Staab (Editor), Rudi Studer

(Editor), Springer 2004

2. Gomez-Perez, A., Corcho. O., Fernandez-Lopez, M.: Ontological Engineering : with

examples from the areas of Knowledge Management, e-Commerce and the Semantic Web.

Asuncion Gomez-Perez, Oscar Corcho, Mariano Fernandez-Lopez, Springer 2004

3. Roche, C.: Ontology: a Survey. 8th Symposium on Automated Systems Based on Human

Skill and Knowledge IFAC, September 22-24 2003, Göteborg, Sweden

4. Buitelaar, P., Cimiano P., Magnini B.: Ontology Learning from Text: Methods, Evaluation

and Applications (Frontiers in Artificial Intelligence and Applications, Vol. 123). P.

Buitelaar (Editor) Ios Press Publication (July 1, 2005)

5. Dourgnon-Hanoune, A., Salaün, P., Roche, C.: Ontology for long-term knowledge. XIX

IEA/AIE, Annecy 27-30 June 2006

6. Dourgnon-Hanoune, A., Mercier-Laurent, E., Roche, C.: How to value and transmit nuclear

industry long term knowledge. ICEIS 2005, 7th International Conference on Enterprise

Information Systems, Miami, 24-28 May 2005

7. Sinclair, J.: Corpus and Text - Basic Principles. In: Developing Linguistic Corpora: a Guide

to Good Practice. Ed. M. Wynne. Oxford: Oxbow Books: 1-16. Available online from

http://ahds.ac.uk/linguistic-corpora/ [Accessed 2007-04-11]

8. Aussenac-Gilles, N., Sörgel, D.: Text analysis for ontology and terminology engineering.

Applied Ontology. n°1. pp. 35-46

9. Daille, B.: Recent Trends in Computational Terminology. Special issue of Terminology

10:1 (2004). Edited by Béatrice Daille, Kyo Kageura, Hiroshi Nakagawa and Lee-Feng

Chien, Benjamins publishing company

10. Harris, Z.: Mathematical Structures of Language. 1968, reprint 1979. R.E. Krieger

Publishing Company, Inc.

11. http://wordnet.princeton.edu/

12. Kiryakov, A., Popov, B., Terziev, I., Manov, D., Ognyanoff, D.: Semantic Annotation,

Indexing, and Retrieval.Elsevier's Journal of Web Sematics, Vol. 2, Issue (1), 2005

13. Grice, H.P.: Meaning. Philosophical Review n°66. pp 377-88, 1957

14. Cruse, D.A.: Lexical Semantics. Cambridge University Press 1986

15. Ushold, M., Gruninger, M.: Ontologies: Principles, Methods and Applications. Knowledge

Engineering Review, Vol. 11, n° 2, June 1996. Also available from AIAI as AIAI-TR-191

16. Gruber, T.: A Translation Approach to Portable Ontology Specifications. Knowledge

Systems Laboratory September 1992 - Technical Report KSL 92-71 Revised April 1993.

Appeared in Knowledge Acquisition, 5(2):199-220, 199

17. Sapir, E.: Language. An Introduction to the study of speech. Docer Publications, 2004.

Originally published by Harcourt, Brace and Company, 1921)

18. Whorf, B.L.: Language, Thought and Reality. The MIT Press, 1956

19. Baader, F., Calvanese, D., McGuiness, D., Nardi, D., Patel-Schneider, P.: The Description

Logic Handbook. Franz Baader, Diego Calvanese, Deborah McGuinness, Daniele Nardi,

Peter Patel-Schneider, editors. Cambridge University Press, 2003

20. Wright, J. M., Fox, M.S., Adam, D.: 84 "SRL/1.5 Users Manual." Technical

report;Robotics Institute, Carnegie-Mellon University 1984.

21. Woods 75. What’s in a Link: Foundations for Semantic Networks. Representation and

Understanding: Studies in Cognitive Science, 35-82, edited by D.G. Bobrow and A.M.

Collins, New York: Academic Press, 1975.

22. OWL Web Ontology Language: http://www.w3.org/TR/owl-features/

23. Guarino, N., Carrara, M., Giaretta, P.: An Ontology of Meta-Level Categories. of

Knowledge Representation and Reasoning: Proceedings of the Fourth International

Conference (KR94), Morgan Kaufmann, San Mateo, CA.

24. Kaplan, A.: Towards a consistent logical framework for ontological analysis. FOIS’01.

October 17-19, 2001, Ogunquit, USA.

25. Roche, C.: The ‘Specific-Difference’ Principle: a Methodology for Building Consensual

and Coherent Ontologies. IC-AI’2001: Las Vegas, USA, June 25-28 2001

26. Spies, M., Roche, C.: Aristotelian ontologies and OWL modelling. Third International

Workshop on Philosophy and Informatics. Saarbrücken, Germany - May 3-4, 2006