=Paper=
{{Paper
|id=Vol-38/paper-7
|storemode=property
|title=Using Information Extraction Rules for Extending Domain Ontologies
|pdfUrl=https://ceur-ws.org/Vol-38/ijcai01-OL_sintek_etal.pdf
|volume=Vol-38
|dblpUrl=https://dblp.org/rec/conf/ijcai/SintekJEA01
}}
==Using Information Extraction Rules for Extending Domain Ontologies==
https://ceur-ws.org/Vol-38/ijcai01-OL_sintek_etal.pdf
Using Information Extraction Rules for Extending Domain Ontologies
- Position Statement for the IJCAI-2001 Workshop on Ontology Learning -
Michael Sintek, Markus Junker, Ludger van Elst, and Andreas Abecker
German Research Center for Artificial Intelligence (DFKI)
– Knowledge Management Group –
P.O. Box 2080, D-67608 Kaiserslautern, Germany
Phone: +49 631 205 3470, Fax: +49 631 205 3210
e-Mail (sintek,junker,elst,aabecker)@dfki.de
1 Ontologies in the FRODO Project two assumptions lead to two characteristics of our approach:
In the FRODO project [1] we aim at the development of Learning ontological information from text documents
a “Framework for Distributed Organizational Memories” should be a main component of the overall scenario. We
(OMs). We start with the observation that knowledge and set the goal already in [3]. In the meanwhile we sket-
expertise is always heavily distributed in an organization. We ched a method for business-process oriented knowledge
accept the fact that this is not an intermediary, imperfect state modeling in the company, realized as an amalgamation
which should be overcome by a central, ontologically struc- of the CommonKADS [6] and the IDEF5 [7] suites of
tured information system, but rather a natural and meaningful methods [2]. We build upon the Protégé-2000 know-
situation (because during the introduction of OM systems it ledge acquisition and modeling tool [8] which we exten-
is normal to start with small, focussed systems which should ded already by some modules for modeling, reasoning,
interoperate later; because much expertise is better to be crea- and visualization (see [1]). We are currently working on
ted, hold, and maintained locally; or because in the case of an integration of the MindAccess(r) commercial [9] text
interorganizational collaborations or virtual teams a deeper analysis workbench which employs a numberof statisti-
integration of information systems cannot be achieved). cal document feature extraction and document analysis
Hence, a main goal of the FRODO project is to develop functionalities.
a scalable, extensible OM middleware built for easy integra-
tion of new components and linking of collaborating com- In order to cope with the complexity and dynamics of
ponents [2]. FRODO builds upon the KnowMore framework real-world usage scenarios for ontologies in a distribu-
for contextually-aware, ontology-based OMs [3,4], but rela- ted OM, we develop a methodological framework for
xes some constraints of the original model, especially the idea understanding and organizing the roles, responsibilities,
of a centralized OM using one overall set of organizational rights, and obligations of actors constituting an ontology
ontologies. society in a complex, agent-based OM system [10].
Besides the technical provisions for such a distributed, In the IJCAI-01 “Ontology Learning” workshop we would
highly dynamic environment, we lay special emphasis on like to discuss primarily an approach for extending the
considerations and methods which are necessary to realize above statistically-oriented learning techniques towards a
such a scenario in industrial practice. In each industrial en- more knowledge-based one using an ILP (Inductive Logic
vironment, besides the questions of smooth introduction of Programming [11]) algorithm which can use more elabora-
new technology regarding human factors and organizational ted document models and can cope with different sources of
processes, and besides the question of modeling tools and sophisticated background knowledge.
method support for knowledge (in particular ontologies for
structuring OMs or parts of OMs) acquisition, at least two
other factors are of utmost importance: 2 Ontology Learning with Information
One is the predominance of informal, i.e. essentially text- Extraction Rules
based, representations of knowledge. This is not only just a
matter of fact, but really useful, because the cost of formaliza- Figure 1 illustrates the overall idea of building ontologies
tion is often not in the right relation to the potential benefits with learned information extraction rules. We start with:
such that many informal parts of the scenario are economi-
1. An initial, hand-crafted seed ontology of reasonable
cally reasonable [5]. One implication is that also methods for
quality which contains already the relevant types of re-
building formal models must be affordable.
lationships between ontology concepts in the given do-
The other is the fact that ontologies are not a stand-alone
main.
component built once and then remaining untouched, but a li-
ving element in the overall scenario, used for different purpo- 2. An initial set of documents which exemplarily represent
ses, communicating with other system parts, and representing (informally) substantial parts of the knowledge represen-
knowledge about a continuously changing world [10]. These ted formally in the seed ontology.
Figure 1: Overall approach for ontology learning with information extraction rules
Now we assume that similar ontological phenomena—e.g. sages as instances of some relationship. This ap-
the fact that relationship R holds between concept A and con- proach is more work-intensive, but promises faster
cept B—are expressed in the text in similar ways. Consider, learning and more precise results. We employed
e.g., a medical domain where the fact that Disease A can be this approach already successfully in an industrial
treated (this is the Relationship R) with Cure B. Such A-R-B information extraction project [12].
instances of relationship R could, for instance, look like: 2. Employ a pattern learning algorithm to automatically
My headache was cured by medication with Aspirin. construct information extraction rules which abstract
Sue’s headache was addressed with acupuncture. from the specific examples, thus creating general state-
ments which text patterns are an evidence for a certain
Cancer can be treated with chemotherapy. ontological relationship. In the example above, such an
Cancer is often treated with surgery. information extraction rule could have the form:
Our main idea is that, (i) given such texts are available In order to detect an instance of the “Method
which explain the ontological knowledge, and (ii) given these B is a possible Cure for Disease A” relation-
texts are sufficiently similar with respect to the question how ship, search for an instance of the concept Di-
similar factual statements are textually represented, it should sease, look whether there is a synonym of the
be possible: word (stem) “treat” in a distance of at most
1. To take the pairs of (ontological statement, one or more two words, search for the word “with” in a di-
textual representations) as positive examples for the way stance of at most two words, directly followed
how specific ontological statements can be reflected in by an instance of the concept Cure.
texts. There are two possibilities to extract such examp- In order to learn such information extraction rules, we
les: need some prerequisites:
Based on the seed ontology, the system looks up the (a) A sufficiently detailed representation of documents
signature of a certain relation (e.g., R links a Di- (in particular, including word positions, which is
sease with a Cure), searches all occurrences of in- not usual in conventional, vector-based learning
stances of the concept classes Disease and Cure, re- algorithms, WordNet-synsets, and part-of-speech
spectively, within a certain maximum distance, and tagging).
regards these co-occurrences as positive examples (b) A sufficiently powerful representation formalism
for relationship R. This approach presupposes that for extraction patterns.
the seed documents have some “definitional” cha- (c) A learning algorithm which has direct access
racter, like domain specific lexica or textbooks. to background knowledge sources, like the al-
The user goes through the seed documents with a ready available seed ontology containing state-
marker and manually highlights all interesting pas- ments about known concept instances, or like the
WordNet database of lexical knowledge linking 4. Abecker, A., Bernardi, A., Hinkelmann, K., Kühn, O.,
words to their synonyms sets, giving access to sub- and Sintek, M. (2000). Context-Aware, Proactive Deli-
and superclasses of synonym sets, etc. very of Task-Specific Knowledge: The KnowMore Pro-
ject. International Journal on Information System Fron-
In [13,14] we present an ILP-like rule learner specifi- tiers, Kluwer, 2(3/4).
cally adapted to the task of pattern-based text classifica-
tion (which can be solved with the same methods as the 5. Buckingham Shum, S. (1997). Balancing Forma-
information extraction task used in the ontology learning lity with Informality: User-Centred Requirements for
application) which fulfills these requirements. In parti- Knowledge Management Technologies. AIKM’97:
cualar, this rule learner relies on a document represen- AAAI Spring Symposium on Artificial Intelligence
tation in which the order of words is preserved. Thus, in Knowledge Management, Stanford University, Palo
learned text patterns can test on the order and distance Alto, CA. AAAI Press.
of specific words. In [16] it is shown how its imple- 6. Schreiber, G., Akkermans, H., Anjeiwerden, A., de
mentation concepts can be mapped to standard ILP ap- Hoog, R., Shadbolt, N., van de Velde, W., and Wielinga,
proaches, which shows how its expressive power with B. (1999). Knowledge Engineering and Management:
respect to pattern representation can even be extended The CommonKADS Methodology. MIT Press.
towards full LP formalisms including recursive rules. In
7. Information Integration for Concurrent Engi-
[15] we elaborate a bit on the integration of background
neering (1994). IDEF5 Method Report. URL:
knowledge sources, especially WordNet.
http://www.idef.com/ .
3. Apply these learned information extraction rules to 8. Grosso, W.E., Eriksson, H., Fergerson, R.W., Gennari,
other, new text documents to discover new or not yet J.H., Tu, S.W., and Musen, M.A. (1999). Knowledge
formalized instances of relationship R in the given app- Modeling at the Millennium (The Design and Evolution
lication domain. of Protege-2000). SMI-1999-0801. Stanford Medical
Lab. URL: protege.stanford.edu
3 Status 9. MindAccess product description (2000). Insiders in-
The algorithm described has not yet been implemented and formation management GmbH, Kaiserslautern. URL:
tested. However, all required prerequisites are available as http://www.im-insiders.de/html/infomaterial.html. In
described above and in [13,14,15,16]. Further, we are in con- German.
tact with several application projects (in the nuclear and the 10. van Elst, L. and Abecker, A. (2001). Ontology-Related
chemical industry) in order to get significant test data. A criti- Services in Agent-Based Distributed Information Infra-
cal factor for the success of the approach will be the question structures. Submitted to: SEKE’01, The Thirteenth
of how typical the textual representations of specific (kinds International Conference on Software Engineering &
of) statements will be in the seed documents. Knowledge Engineering, June 13-15, 2001, Buenos Ai-
Compared to other ontology learning approaches it should res - Argentina
be noted that our technique is not restricted to learning ta- 11. Lavrac, N. and Dzeroski, S. (1994). Inductive Logic
xonomic relationships, but arbitrary relationships in an app- Programming: Techniques and Applications. Chiches-
lication domain. We expect that, in contrast to more statisti- ter, UK: Ellis Horwood.
cally oriented approaches, which tend to result in too many
candidate results (because of many possibly relevant word 12. ANNOCLASS project description. URL:
co-occurences), our approach needs more input and assumes http://www.dfki.de/pas/f2w.cgi?daimc/annoclass-e
more prerequisites, but found relationship candidates will be 13. Junker, M. (2000). Heuristisches Lernen von Regeln für
correct with a higher probability. die Textkategorisierung. Dissertation. Fachbereich In-
formatik. Universität Kaiserslautern. In German.
References 14. Junker, M. and Abecker, A. (1998). Learning Com-
plex Pattern for Document Categorization. In: AAAI-
1. FRODO project homepage: http://www.dfki.uni- 98/ICML Workshop on Learning for Text Categoriza-
kl.de/frodo/ tion. Madison, Wisconsin, USA.
2. Abecker, A., Bernardi, A., van Elst, L., Lauer, A., Maus, 15. Junker, M. and Abecker, A. (1997). Exploiting Thesau-
H., Schwarz, S., and Sintek, M. (2001). FRODO: A Fra- rus Knowledge in Rule Induction for Text Classification.
mework for Distributed Organizations - Milestone M1: In: RANLP’97 - Recent Advances in NLP, pp. 202-207,
Requirements Analysis and System Architecture. DFKI Tzigov Chark, Bulgaria.
Document D-01-01. In preparation. Partially in Ger-
man. 16. Junker, M., Sintek, M., and Rinck, M. (2000). Learning
for Text Categorization and Information Extraction with
3. Abecker, A., Bernardi, A., Hinkelmann, K., Kühn, O., ILP. In Learning Language in Logic, Springer, LNCS
and Sintek, M. (1998). Towards a Technology for Orga- 1925.
nizational Memories. IEEE Intelligent Systems, 13(3),
May/June.