=Paper=
{{Paper
|id=Vol-101/paper-12
|storemode=property
|title=Requirements for Information Extraction for Knowledge Management
|pdfUrl=https://ceur-ws.org/Vol-101/Philipp_Cimiano-et-al.pdf
|volume=Vol-101
}}
==Requirements for Information Extraction for Knowledge Management==
https://ceur-ws.org/Vol-101/Philipp_Cimiano-et-al.pdf
Requirements for Information Extraction
for Knowledge Management
∅ ⊕
Philipp Cimiano*, Fabio Ciravegna , John Domingue ,
∅
Siegfried Handschuh*, Alberto Lavelli+, Steffen Staab*, Mark Stevenson
*AIFB, University of Karlsruhe
∅ NLP Group, University of Sheffield
⊕ Open University, Milton-Keynes
+ ITC-irst, Trento
ABSTRACT readable data, i.e. metadata.
Both scenarios are comparable to the extent mentioned
Knowledge Management (KM) systems inherently suffer above and in fact similar solutions have been proposed to
from the knowledge acquisition bottleneck - the difficulty overcome part of the problems associated with them. On the
of modeling and formalizing knowledge relevant for one hand, ontologies have been proposed as a formalism to
specific domains. A potential solution to this problem is externalize and share knowledge within KM [Staab et al.
Information Extraction (IE) technology. However, IE was 02, Fensel 01, Mulholland et al. 01, Benjamins 98] as well
originally developed for database population and there is a as in the context of the Semantic Web [Berners-Lee et al.
mismatch between what is required to successfully perform 01]. Ontologies are suitable for this purpose because they
KM and what current IE technology provides. In this paper represent a formal, explicit specification of a shared
we begin to address this issue by outlining requirements for conceptualization [Gruber 93]. A shared conceptualization
IE based KM. in this sense has to be understood as an abstract model of
Keywords: Information Extraction, Knowledge Management, some aspect or part of the world shared by a certain group
Ontologies, Annotation of people with a common interest. Formal and explicit refer
to the fact that such an ontology should also be readable for
machines. On the other hand, semi-automatic or automatic
1 INTRODUCTION
methods have been proposed for KM as well as for the
Semantic Web in order to reduce the cost of producing
A large part of a company’s knowledge is stored in textual metadata [Ciravegna et al. 02], [Handschuh et al. 02]
documents available within intranets. However, this [Vargas-Vera et al. 02].
knowledge cannot be queried nor captured in a
straightforward way, which reduces a company’s In this context, Information Extraction from text (IE) is a
efficiency. The challenge is to formally represent the very promising technique for the Semantic Web as well as
knowledge contained in textual form such that it can be for KM [Ciravegna 01]. IE is an automatic method with the
accessed and used by the workers in an enterprise through purpose of locating relevant entities and facts in electronic
various knowledge-based services. documents for further use and fits perfectly into the KM
scenario described above. A first requirement derived from
A similar scenario is encountered within the Semantic Web this potential use of IE within KM is the fact that the target
in which the central idea is to provide efficient access to knowledge structures produced by the IE system have to be
heterogeneous and distributed web resources. This is only compatible with the ontology used for formalizing
possible if the knowledge contained in the resources has externalized knowledge. Only then can the extracted
been formalized so that it can be shared, understood and knowledge be shared and further processed within a
reused by other people or applications, such as crawlers, company’s KM environment. This paper focuses on the
information brokering services and mediators. So the way IE could be integrated into the existing KM technology
success clearly depends on the availability of machine- as well as on the requirements that such integration poses
on the IE and KM technologies. regard to the ontology so that they can be fed back into the
enterprise’s archive for further use. This is where IE
The remainder of this paper is organised as follows: in
techniques come into play. As mentioned earlier, IE can be
Section 2 we discuss the requirements for the integration of
applied either in an automatic or semi-automatic way in
IE into Knowledge Management Systems. The
order to produce annotations which are consistent with a
requirements such integration poses on the IE technology
given ontology. Thus, IE should directly exploit the
itself are then covered in Section 3. The paper finishes with
underlying ontology in order to produce compatible
some conclusions and implications.
knowledge structures. In particular, the mapping from
knowledge structures produced by IE to ontological models
represented in languages such as DAML+OIL, RDF(S) or
2 Knowledge Management Requirements OWL should be straightforward. An issue related to this
requirement is the necessity to produce relational metadata,
The most important requirement for a KM solution is its - instances of relations defined in the selected ontology.
successful integration into the enterprise in question. The One further important requirement is the need for some
process concerned with the introduction of a KM system as quality control of the output produced by IE before further
well as its maintenance, evolution and refinement is processing it for KM purposes. In fact, IE is by definition
commonly referred to as the knowledge meta process an error-prone process. Consequently, the resulting
[Staab et al. 02]. The knowledge process on the other hand knowledge structures cannot be directly used to populate an
is concerned with issues related to the use of the introduced ontology without manual intervention. This quality control
KM solution. In particular, it focuses on the cycle of can for example take place directly in an annotation tool
information creation, capture, retrieval and use, for example integrating IE as a plug-in. In this sense, the annotation
to create new information and close the cycle (see [Staab et framework would thus suggest annotations to the user,
al. 02]) It is important that this cycle fits with existing (and which have to be manually validated. We will make use of
emerging) work practices. Both processes are dependent on OntoMat Annotizer [Handschuh et al. 02] or MnM [Vargas-
each other as the refinement of the KM solution can only Vera et al. 02] for this purpose. However, it could also be
take place by considering the working knowledge process, thought of having an ‘on the fly’ validation of produced
which in turn will be modified according to the introduced annotations in the sense that users may decide at some point
refinements. The information obtained in the retrieval/ during their work if a specific annotation is plausible or not
access step of the knowledge management cycle is then and thus whether it can be kept or has to be rejected.
typically included within a specific application and can also
be used in the creation of new documents (see Figure 1). Documents are created in a context that is not captured in
the text. It is thus important that annotations not only
reflect the explicit content of a particular document but also
Ontology
Creation knowledge related to its creation context, for example,
Enterprise Archive
Information
reasons why particular items were omitted. Nevertheless,
Ontology (WR,BP)
Capture such an annotation should also be consistent with the
IE/Annotation
IE/Annotation underlying ontological model used within the enterprise so
that this knowledge can be stored and reused as with
Evolution
‘conventional’ document annotations.
Inference
Engine
Document
Authoring
Finally, it is important to mention that it cannot be expected
that a reasonable and suitable ontology will be available
New WR right from the beginning. Moreover, we envision starting
Metadata Document
Queries Search from a small seed ontology, which will be constantly
Document
Information
Retrieval extended, refined and modified. We intend to create such a
Browsing
seed ontology with the help of the text mining approach
Document
Clustering
Clustering
Information presented in [Cimiano et al. 03]. Thus the knowledge
Creation
process and the knowledge meta process [Staab et al. 02]
will be highly interleaved and dependent on each other. In
this context it is important that knowledge about changes in
the ontology is also made explicit and to have some
Figure 1 ontology evolution support such as described in [Stojanovic
et al. 02].
In order to close the knowledge process cycle, the 3 Information Extraction Requirements
information contained in the newly created documents has
to be captured, i.e. the documents have to be annotated with
Research in IE has been largely driven by the Message types into single system has been done in Amilcare
Understanding Conferences (MUC). These competitions [Ciravegna 03]. This anyway still represents a challenge to
focused on extracting information from newswire text. The the language processing community [Ciravegna 01].
participants were required to perform different tasks, from
the identification of person, location and organization
names (Named Entity recognition) to the identification of 3.3 Adaptivity and Usability
relations between entities (Template Relation) to the
construction of complex templates (Scenario Template). Traditional IE systems have tended to be difficult to port to
The original aim of IE was to automatically fill database new domains and extraction tasks. For example, [Lehnert
records from text and consequently systems have not, in et. al. 92] estimated that 1,500 person-hours of highly
general, been designed to carry out knowledge markup. In skilled labor were required to adapt their system for MUC-
the remainder of this section we discuss the requirements 4. Clearly the applications will be limited for any tool that
for IE systems performing knowledge markup in the context requires such an investment to be adapted to a new domain
of KM. or extraction task.
It is therefore vital that IE systems can be adapted with the
3.1 Relation Extraction least possible effort and that this process can be carried out
by non-experts. Machine learning techniques could be used
for this (e.g [Soderland 99], [Yangarber et. al. 00,
[Handschuh et al. 02] discuss the problems involved in Yangarber 03]). Interaction with annotation tools requires
using an IE system which carries out concept recognition little more than marking relevant concepts in text. However,
(e.g. Amilcare [Ciravegna 03]) to produce relational the mode of interaction for marking relations in text is not
metadata, i.e. instances of a certain ontological relation. For as obvious as for marking concepts, which can be directly
example, in the sentence “Mr. Jones was hired by Dot.Kom highlighted.
Ltd. last week” Amilcare can identify “Mr. Jones” as a
person (and even as a “hiredPerson”) and “Dot.Kom Ltd.” In general, the IE systems must be portable by non experts
as a company (or even “hiringCompany”). However, it and users should be assisted in the whole application
cannot identify the relation between these two entities (i.e., lifecycle. [Ciravegna 01] identifies the requirements in this
that the specific person was hired by the specific company; respect, mentioning the need for tools for (1) scenario
this means that if different hiringCompany and hiredPerson definition, (2) system adaptation and result validation and
exist it is not possible to connect them properly). (3) application delivery. Scenario design is not an issue in
[Handschuh et al..02] present a discourse analysis approach ontology-based IE because the ontology will provide the
to map the entities tagged by Amilcare into graph structures scenario. Concerning system adaptation and result
such as those used in ontological formalisms as RDF, validation, experiences such as Melita [Ciravegna et al. 02]
DAML+OIL or OWL. In order to use an IE system for KM show that a great deal of control can be reached using
purposes it is necessary that it produces relational metadata simple HCI techniques. We are currently investigating in
that can be used to directly populate an ontology. This the direction of further improvement of usability through
means that some form of relation extraction is necessary strong integration with the ontology as explained below.
(e.g. [Soderland 99, Yangarber et al. 00, Yangarber 03]).
3.4 Interaction with ontologies
Such a component could be trained on relational annotation
produced by a system like the OntoMat Annotizer
[Handschuh et al. 02]. This type of approach could be It is crucial for the integration of IE into KM that its output
supplemented by an ontology-based discourse analysis can be directly used to populate ontologies or to enrich
approach such as the one proposed in [Handschuh et al. documents with ontology-based metadata. Thus, it is
02]. important that the output of IE systems can be mapped in a
straightforward way to ontological models coded in
languages such as RDF(S), DAML+OIL or OWL.
3.2 Text types processed Essentially this has four implications for IE:
The systems that participated in the MUC evaluations were 1) Detecting concepts over a hierarchy: IE should
required to extract information from well-formed newswire directly interact with the ontological hierarchy and
text. However, a KM system should be able to process a tag instances at different levels of hierarchical
wider variety of texts since they will be expected to process abstraction. From a practical point of view rules
web and intranet pages. IE systems have tended to extract should be generalized semantically using the
information from a limited variety of text types, for ontology.
example free and semi-structured text [Soderland 99] or
tabular data [Hurst 00]. Initial attempt to cover all these
2) Exploiting conceptual markup as context: It is In summary, the successful integration of IE into KM
possible to imagine that IE systems could operate methodology presupposes a strong and direct interaction
in a bootstrapping-like fashion and make use of between the ontology, the IE system as well as the
conceptual markup to extract the conceptual constantly changing information needs of the users.
relation between two previously tagged entities.
3) Exploiting lexical information: It would be useful
ACKNOWLEDGMENTS
to include information about how certain
This work was carried out within the IST-Dot.Kom project
conceptual relations are expressed linguistically in
(http://www.dot-kom.org), sponsored by the European
a text. This could for example allow the rule
Commission as part of the framework V, (grant IST-2001-
induction algorithm a more efficient exploration of
34038). Dot.Kom involves the University of Sheffield
the search space. Information about synonyms
(UK), ITC-Irst (I), Ontoprise (D), the Open University
such as contained in linguistic ontologies as
(UK), Quinary (I) and the University of Karlsruhe (D) . Its
WordNet [Miller 90] could also turn out very
objectives are to develop Knowledge Management and
useful in the context of the semantic generalization
Semantic Web methodologies based on Adaptive
of extraction rules (e.g. [Chai et. al. 99] and
Information Extraction from Text.
[Harabagiu et al. 00]).
4) Mapping between tags and concepts: The
mapping between the IE system and ontology REFERENCES
should be one-to-one to allow the ontology to be [Berners-Lee et al. 01] Berners-Lee, T., Hendler, J. and Lassila, O. “The
exploited within the IE system and use the Semantic Web”. Scientific American, May 2001 Issue.
annotation produced by the IE system to populate [Benjamins et al. 98] Benjamins, V., Fensel, D., Gómez Pérez, A.
the ontology. “Knowledge Management through Ontologies”. In Proceedings of
the Conference on Practical Aspects of Knowledge Management
(PAKM), 1998.
Summarizing, the above mentioned requirements suggest
[Chai et. al. 99] Chai, J., Biermann, A. and Guinn, C. “Two Dimensional
some relevant directions for improving IE so that it can Generalization in Information Extraction” In Proceedings of the
successfully satisfy KM needs. First of all, the importance Sixteenth National Conference on Artificial Intelligence (AAAI-99)
of relation extraction will be addressed further investigating 1999.
the approaches described in [Yangarber et al. 00,
[Cimiano et al. 03] Cimiano, P., Staab, S. and Tane, J. “Automatic
Yangarber 03]. Such unsupervised approaches take into Acquisition of Taxonomies: FCA meets NLP”. In Proceedings of the
consideration also the issue of adaptivity, crucial for ECML/PKKD Workshop on Adaptive Text Extraction and Mining,
reducing the cost of porting to new domains and 2003.
applications. Adaptivity will be dealt with also [Ciravegna 01] Ciravegna, F. “Challenges in Information Extraction from
experimenting with bootstrapping techniques, such as co- Text for Knowledge Management”, In IEEE Intelligent Systems and
training. Their Applications (Trends and Controversies), 2001.
[Ciravegna et al. 02] Fabio Ciravegna, Alexiei Dingli, Daniela Petrelli and
Yorick Wilks: User-System Cooperation in Document Annotation
4 Summary and Conclusion based on Information Extraction” in Proceedings of the 13th
International Conference on Knowledge Engineering and
Knowledge Management (EKAW 2002), 1-4 October 2002 -
In the context of KM, IE cannot be regarded as a stand- Sigüenza (Spain).
alone tool which can be applied quite independently of the
KM technology used. In fact, it is important for the IE [Ciravegna 03] Ciravegna, F. “Designing Adaptive Information Extraction
system to directly interact with the ontology to extract for the Semantic Web in Amilcare” in S. Handschuh and S. Staab
(eds), Annotation for the Semantic Web IOS Press, Amsterdam,
knowledge which is compatible with it and can thus be 2003.
reused within the enterprise`s KM environment.
Furthermore, the information extraction system should [Fensel 01] Fensel, D. Ontologies: A Silver Bullet for Knowledge
certainly be adaptive and applicable to a wide range of text Management, Springer Verlag, 2001.
types and genres. Concerning the knowledge cycle, it seems [Gruber 93] Gruber, T. “A translation approach to portable ontology
very important that the meta knowledge process and the specifications”. Knowledge Acquisition, 5:199-220, 1993.
knowledge process are highly interleaved and that the user
[Handschuh et al. 02] Handschuh, S., Staab, S. and Ciravegna, F. “S-
is supported in the meta knowledge process by (semi-) CREAM - Semi-automatic CREAtion of Metadata”, In Proceedings
automatic methods to produce a seed ontology which will of the 13th International Conference on Knowledge Engineering
be iteratively refined according to requirements derived and Knowledge Management (EKAW02), 2002.
from the working knowledge process.
[Harabagiu et al. 00] Harabagiu, S. and Maiorano, S. “Acquisition of [Stanoevska et al. 98] Stanoevska, K., Hombrecher, A., Handschuh, S.
Linguistic Patterns for Knowledge-Based Information Extraction” In and Schmid, B. “Efficient Information Retrieval: Tools for
Proceedings of the Language Resources and Evaluation Knowledge Management”. In Proceedings of the Conference on
Conference (LREC-2000), Athens, Greece, 2000. Practical Aspects of Knowledge Management, 1998.
[Hirschman 1998] Hirschman, L. "The Evolution of evaluation: Lessons [Stojanovic et al. 02] Stojanovic, L., Stojanovic, N. and Maedche, A..
from the Message Understanding Conferences", Computer Speech “Change discovery in ontology-based knowledge management
and Language, 12, pp. 281-305, 1998. systems”. In Proceedings of the 2nd International Workshop on
Evolution and Change in Data Management (ECDM), 2002.
[Hurst 2000] Hurst, M. “Processing Tables for Information Extraction”
Ph.D. thesis, University of Edinburgh. [Yangarber et al. 00] Yangarber, R., Grishman, R., Tapanainen, P. and
Huttunen, S. “Automatic Acquisition of Domain Knowledge for
[Lehnert et. al. 1992] Lehnert, W. and Cardie, C. and Fisher, D. and Information Extraction”. In Proceedings of the 18th International
McCarthy, J. and Riloff, E. and Soderland, S. “University of Conference on Computational Linguistics (COLING), 2000.
Massachusetts: Description of the CIRCUS System used for MUC-
4” Proceedings of the Fourth Message Understanding Conference [Yangarber 03] Yangarber, R. “Counter-Training in Discovery of
(MUC-4) Semantic Patterns”. In Proceedings of the 41st Annual Meeting of
the Association for Computational Linguistics (ACL 2003), 2003.
[Miller 90] Miller, G. “WordNet: An On-line Lexical Database”
International Journal of Lexicography 3(4):235-312 [Vargas-Vera et al. 02] Vargas-Vera, M., Motta, E., Domingue, J.,
[Mulholland et al. 01] Mulholland, P., Zdrahal, Z.,Domingue, J., Hatala, Lanzoni, M., Stutt, A. and Ciravegna, F. “MnM: Ontology Driven
M. and Bernardi, A. A Methodological Approach to Supporting Semi-Automatic and Automatic Support for Semantic Markup”. In
Organisational Learning. International Journal of Human Proceedings of the 13th International Conference on Knowledge
Computer Studies. Vol. 55, No. 3, September 1, 2001, pp. 337-367 Engineering and Management (EKAW 2002), ed. Gomez-Perez, A.,
Springer Verlag, 2002.
[Soderland 99] .Soderland, S. “Learning Information Extraction Rules for
Semi-Structured and Free Text”. Machine Learning 34(1-3), 1999.
[Staab et al. 02] Staab, S., Studer, R. and Sure, Y. “Knowledge Processes
and Meta Processes in Ontology-based Knowledge Management”. In
Handbook of Knowledge Management, Springer Verlag, 2002.