=Paper=
{{Paper
|id=Vol-99/paper-2
|storemode=property
|title=Knowledge Extraction by Using an Ontology Based Annotation Tool
|pdfUrl=https://ceur-ws.org/Vol-99/Maria_Vargas-Vera-et-al.pdf
|volume=Vol-99
|dblpUrl=https://dblp.org/rec/conf/kcap/Vargas-VeraMDSL01
}}
==Knowledge Extraction by Using an Ontology Based Annotation Tool==
https://ceur-ws.org/Vol-99/Maria_Vargas-Vera-et-al.pdf
Knowledge Extraction by using an Ontology-based
Annotation Tool
Maria Vargas-Vera, Enrico Motta, John Domingue,
Simon Buckingham Shum and Mattia Lanzoni
Knowledge Media Institute (KMi),
The Open University,
Walton Hall, Milton Keynes, MK7 6AA, United Kingdom
m.vargas-vera, e.motta, j.b.domingue, s.buckingham.shum, m.lanzoni @open.ac.uk
ABSTRACT from texts by using a machine learning component called Crystal.
This paper describes a Semantic Annotation Tool for extraction of To extract rules from text, we had developed an environment which
knowledge structures from web pages through the use of simple allows user to perform four phases: browse, semantic annotation
user-defined knowledge extraction patterns. The semantic annota- of pages, learning rules and information extraction (IE) of the web
tion tool contains: an ontology-based mark-up component which pages. Each of these phases are described as follows:
allows the user to browse and to mark-up relevant pieces of in-
1. Browse
formation; a learning component (Crystal from the University of
Massachusetts at Amherst) which learns rules from examples and This option could by used by the user to select the kind of
an information extraction component which extracts the objects and browser in our case could be WebOnto or any other browser.
relation between these objects. Our final aim is to provide support WebOnto [3] provides web-based visualisation, browsing and
for ontology population by using the information extraction com- editing support for the ontology. It allows easier develop-
ponent. Our system uses as domain of study “KMi Planet”, a Web- ment and maintenance of the knowledge models, themselves
based news server that helps to communicate relevant information specified in OCML (Conceptual Modeling Language) [8].
between members in our institute.
2. The Markup phase. The activity of semantic tagging refers
to the activity of annotating text documents (written in plain
Keywords ASCII or HTML format) with an tags set defined on the
Ontology-based mark-up, Ontology population, Extraction of knowl- ontology, in particular we work with the hand-crafted KMi
edge, Information extraction technologies. ontology (ontology describing Knowledge Media Institute).
The semantic annotation tool provides means to browse the
event hierarchy (described in next section). In this hierarchy
1. INTRODUCTION each event is a class and the annotation component extracts
Semantic annotation has been focused in isolated annotations of the set of possible tags from the slots defined in each class.
web pages. However, semantic web tries to achieve the annota- In general mark-up process might be difficult but in our case
tion of pages with semantic information. In other words, the aim the annotation component is guiding the user with the possi-
is to enrich the content of web pages. Recent work on semantic ble entities which could be marked in the text.
annotation guided by an ontology is discussed in [14]. However, Other approach related to our work is the SHOE Knowledge
our approach has a different aim, we use the ontology as guider to annotator which is a Java program that allows users to mark-
the human annotator of the training set (ie. the user is presented up web pages with the SHOE ontology [5]. However, in
with a set of possibles tags which could be used during the mark- SHOE there is not relation between the new annotations and
up process), and then the system learns rules by using the semantic the original text.
annotations, whilst in OntoAnnotate [14] the user selects the object
identifier and the appropriate class for it from a hierarchy of classes. 3. Learning phase. This phase uses the marked text as training
Then all the information which is in the Ontology for that particular set and learns relations from the stories. It uses crystal as a
object identifier is presented to the user. If the object identifier is learning component. Crystal works using the bottom-up ap-
not defined the user could create a new object or class relation. proach. It finds rules for specific instances and it generalises
One target of the system presented in this paper is to learn rules these rules.
4. The information extraction phase. The goal of a Infor-
mation Extraction system (IE) is to extract specific types of
information from text. For example, an IE system in the
Permission to make digital or hard copies of all or part of this work for domain of KMi (Knowledge Media Institute) organisation,
personal or classroom use is granted without fee provided that copies are should be able to extract the name of KMi projects, KMi
not made or distributed for profit or commercial advantage and that copies funding organisations, awards, dates, etc. The main advan-
bear this notice and the full citation on the first page. To copy otherwise, to tage of IE task is that portions of a text that are not relevant
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee. to the domain can be ignored. Therefore text could be pro-
Copyright 2000 ACM 0-89791-88-6/97/05 ..$5.00 cessed quickly.
Most IE systems use some form of partial parsing to recog-
nise syntactic constructs without generating a complete parse
tree for each sentence. Such partial parsing has the advan-
tages of greater speed and robustness. High speed is neces-
sary to apply the IE to a large set of documents.
IE has been used in several domains, for instance, scientific
articles such as MEDLINE (it contains abstracts of biomedi-
cal journals) [2], bibliographic notices [9], and medical records
[13]. Also, ontologies has been used in IE systems to help
them extract relations from semi or unstructured documents,
statements or terms [11]. Recent work on semi-automatic
ontology acquisition by means of IE, supported by machine-
learning methods, is described in [6, 4]. In similar lines there
is the CMU’s approach for extracting information from hy-
pertext using machine learning techniques (Bayes classifier)
and making use of an ontology [1]. However, we remark that
we are not creating an ontology, we are supporting ontology
population. The ontology population problem is an impor-
tant issue to be addressed since it is difficult to keep up to
date a hand-crafted ontology.
In our work, we had integrated the hand-crafted KMi On-
tology into the information extractor. The main task of the
ontology is to disambiguation of some extracted informa-
tion. For instance, in the event conferring an award “X was
granted Y amount of money”. X could be instantiated to
name of project or institution. In this case we make use of
the ontology to clarify the type of X. Figure 1: System overview
In the construction of our IE component we had integrated
several components (Marmot, Badger and Crystal) from the 2. EVENT TYPOLOGY
University of Massachusetts at Amherst (UMass) which are KMi ontology consists of KMi projects, people in KMi, events,
fully described in Rillof[10]. We remark that in our IE com- etc. In particular we will focus in a section of the KMi ontology
ponent the template matching itself is supported semanti- called events (activities happening in our Institute). The events are
cally by referring to the ontology, but also contains some defined formally in our ontology as classes. Currently, in our KMi
lightweight NLP techniques in order to syntactically identify ontology we have defined 40 different types of events. As the event
some fragments of the sentences. We believe it is important typology is already defined in the KMi ontology. Then, for each
to mix the syntactic and semantic. The semantic checking is event we already had defined the slots which might be instantiated
often necessary to resolve ambiguities, for example, ontolo- by the IE component. Figure 2 shows a portion of the hierarchy of
gies can provide us with axioms of common sense knowl- events as defined in KMi ontology.
edge such “if someone is visiting a place then this someone For the sake of space, we only present the structure of three
should be a person.” Conversely, some grammar construc- type of events from the event hierarchy: visiting-a-place-or-people,
tions (such as dates) can be recognized robustly. conferring-a-monetary-award and demonstration-of-technology.
Figure 1 illustrates the four phases. In particular la browse phase Class Event 1: visiting-a-place-or-people
has been launched.
Our primary contribution is to integrate a template-driven IE en- slots:
gine with an ontology engine (including inference capabilities be-
sides lexicons such as Wordnet) in order to supply the necessary visitor (list of person(s))
semantic content and then to disambiguate extracted information people-or-organisation-being-visited
and finally our second contribution is to provide support for the (list of person(s) or organization)
ontology population process. has-duration (duration)
The paper is organised as follows: In Section 2 we present a ty- start-time (time-point)
pology of two events as are defined in KMi ontology. Section 3 end-time (time-point)
presents the mark-up phase. Section 4 shows the learning phase has-location (a place)
using Crystal. Section 5 presents the extraction of information us- other agents-involved (list of person(s))
ing Badger. Section 6 describes the use of ontology to cope with main-agent (list of person(s))
the ambiguity in the identification of objects in the story. Section 7
shows the OCML 1 code generated after badger obtains template The structure of Event 1 (visiting-a-place-or-people) describes a
instantiations. Section 8 discusses the process of populating an on- set of objects which might be encountered in story describing an
tology as an activity in the life cycle of the ontology construction. event visit, such as, visitor, people-or-organisation-being-visited,
Finally, Section 9 gives conclusions and directions for future work. other agents-involved, etc.
OCML is a language designed for knowledge modeling Class Event 2: conferring-a-monetary-award
other agents-involved (list of person(s))
main-agent (list of person(s))
location-at-start (a place)
location-at-end (a place)
medium-used (equipment)
subject-of-the-demo (title)
Event 3 contains the structure for the event “demonstration-of-
technology”. Entities that need to be recognised are technology,
place, etc.
3. MARK-UP PHASE
The mark-up component aims to help the manual annotation of
web pages. In this component the ontology plays a important role
guiding the mark-up process. The user does not know which is the
relevant information which might be annotated. Therefore, we con-
sider that is useful to have a such tool that presents user with pos-
sibles tags. An example of annotated story is shown in Figure 4.
The user selects an specific class on the hierarchy of events, for
example, “visiting-a-place-or-people”. Then a set of possibles tags
is presented to the user for the event “visiting-a-place-or-people”.
The set of tags are: has-duration, start-time, end-time, has-location,
other agents-involved, main-agent, visitor, people-or-organisation-
being-visited. From this set the user could select a subset of tags
and then automatically a template for the event “visiting-a-place-
or-people” is created. The created template is used later by the
component which make instantiations of templates (Badger). Fig-
ure 3 shows the user selection. In this particular example the user
only selects start-time, end-time, has-location and visitor.
Figure 2: Event hierarchy
slots:
monetary award (sum of money)
has-duration (duration)
start-time (time-point)
end-time (time-point)
has-location (a place)
main-agent (list of person(s))
other agents-involved (list of person(s))
location-at-start (a place)
location-at-end (a place)
awarding-body (an organization)
has-award-rationale (project goals)
In the event 2 the value for the slot has-award-rationale is ex-
tracted from text by using heuristics such as if the word goal ap- Figure 3: Selection of tags
pears in the story then the system will extract as rationale the sen-
tence until it finds full stop. The reason for this is because is to For the sake of space, let us assume that the user annotates the
general to be learned by an IE component. It does not follow any story with two tags: visitor and place from the selected set. Figure 4
grammar rule about how the rationale could be expressed by a jour- shows the semantic annotations which automatically are inserted in
nalist who writes an story describing a project’s award. the text. In the story David Brown was annotated as visitor and
The OU is annotated as place.
Class Event 3: demonstration-of-technology
technology-being-demostrated (technology) 4. LEARNING PHASE
has-duration (duration) This phase was implemented by integrating two tools Marmot
start-time (time-point) and the learning component called Crystal both from Umass.
end-time (time-point) A brief description of Marmot (a text preprocessor) is giving
has-location (a place) before the learning component Crystal is presented.
Figure 4: Annotated story
4.1 Marmot of THE OU and finally a punctuation symbol (PUNC) is the full
Marmot (from UMass) is a natural language preprocessing tool stop is encountered at the end of the sentence.
that accepts ASCII files and produces an intermediate level of text SUBJ(1): DAVID BROWN %COMMA% UNIVERSITY
analysis that is useful for IE applications. Sentences are sepa- PP (2): FOR INDUSTRY
rated and segmented into noun phrases, verb phrases prepositional VB (3): VISITS
phrases. OBJ1(4): THE OU
Marmot has several functionalities: preprocesses abbreviations PUNC(5): %PERIOD%
to guide sentence segmentation, resolves sentences boundaries, iden-
tifies parenthetical expressions, recognises entries from a phrasal In the same fashion, in sentence number 3, DAVID BROWN is
lexicon and replace them, recognises dates and duration phrases, recognised as subject, the word VISITED is recognised as verb and
performs phrasal bracketing of noun, preposition and adverbial phrases, OBJ1 as THE OU.
finally scopes conjunctions and disjunctions.
SUBJ(1): DAVID BROWN %COMMA% THE CHAIRMAN OF
We had defined our own verbs, nouns, abbreviations and tags in
THE UNIVERSITY
order to apply Marmot to our KMi domain. For the sake of space
PP (2): FOR INDUSTRY DESIGN AND IMPLEMENTATION
we would analyse only the first three sentences in the story given
ADVISORY GROUP AND CHAIRMAN OF MOTOROLA
in Figure 5.
PUNC(3): %COMMA%
In the first sentence, Marmot recognised two entities firstly a
VB (4): VISITED
subject (SUBJ) which is JOHN DOMINGUE and secondly a date.
OBJ1(5): THE OU
The latest is recognised and marked between the symbol “@”. Dates
are recognised robustly as regular expressions. 4.2 Crystal
Crystal is a dictionary induction tool. It derives a dictionary of
SUBJ(1): JOHN DOMINGUE concept node (CN) from a training corpus. The first step in dictio-
ADVP(2): @WED_%COMMA%_15_OCT_1997@ nary creation is the annotation of a set of training texts by a domain
PUNC(3): %PERIOD% expert. Each phrase that contains information to be extracted is
tagged (with SGML style tags).
In sentence number 2, DAVID BROWN is recognised as sub- Crystal initialises a CN dictionary for each positive instance of
ject (SUBJ), a prepositional phrase (PP) “FOR INDUSTRY” is en- each type of event. The initial CN definitions are designed to ex-
counter, the verb (VB) VISITS is also found, OBJ1 takes the value tract the relevant phrases in the training instance that creates them
Figure 5: Marmot output Figure 6: Crystal output
ing terrorist activities in Latin America):
but are too specific to apply to a unseen sentences. The main task of
“A visitor from Colombia was hurt when two terrorists attempted
Crystal is to gradually relax the constraints on the initial definitions
to kill the major”.
and also to merge similar definitions.
if visitor from Colombia is marked as victim two terrorist are
Crystal finds generalisations of its initial CN definitions by com-
marked as perpetrators and major as victim.
paring definitions that are similar. This similarity is deduced by
Crystal generates 3 frame cases that represents the following pat-
counting the number of relaxations required to unify two CN def-
terns:
initions. Then a new definition is created with constraints relaxed.
If a text contains the expression “X was hurt” then the system
Finally the new definition is tested against the training corpus to
extracts “X” as the victim.
insure that it does not extract phrases that were not marked with
If a text contains the expression “X attempted to kill” then the
the original two definitions. This means that Crystal takes similar
system extracts “X” as perpetrator.
instances and generalises into a more general rule by preserving the
If the text contains the expression “attempted to kill Y” then the
properties from each of the CN definitions which are generalised.
system extracts “Y” as the victim.
The inductive concept learning in Crystal is similar to the induc-
In recent years had been great interest in annotated-based tech-
tive learning algorithm described in [7] a specific-to-general data-
niques for producing automatically dictionaries. The reason for this
driven search to find the most specific generalisation that covers
is that automatic creation of conceptual dictionaries is important
all positive instances. Crystal finds the most specific generalisation
factor for portability and scalability of an IE system.
that covers all positive instances but uses a greedy unification of
Crystal has been tested on corpus of 300 KMi stories. Crystal
similar instances rather than breadth-first search.
was able to induce a dictionary of CN definitions for each event in
Coming back to our example David Brown’s story. We have that
KMi ontology.
Crystal learns a conceptual node such as the one shown in Figure 7.
These conceptual node states that “X visited”. So that in the
future whenever the pattern “X visited ” appears in the text the case 5. EXTRACTION PHASE
frame will extract “X” as the visitor. A third component called Badger (from UMass) which was also
For the pattern X visited Y, we basically are extracting relations integrated into our IE component.
r(X,Y) from texts which could be interpreted as “X visited Y” and Badger makes the instantiation of templates. The main task of
the Lexicon for relation r is the union of the lexicon(X) and lexi- badger is to take each sentence in the text (in our case a story writ-
con(Y). If we find this relation in our texts then we find a instance ten in a e-mail message) and see if it matches any of our CN defini-
for the event “visiting-a-place-or-people”. tions. If no extraction CN definition applies to a sentence, then no
In this example we do not have the case that two different tem- information will be extracted; this means that irrelevant text can be
plates might apply to the same sentence. But it is possible to en- processed very quickly.
counter these cases. Let us consider the following example from It might occurs that Badger obtains more than one type of event
the MUC domain (the MUC domain is a set of documents describ- for an story. Then our IE system decides to classify the story ac-
Verb: visited (active verb)
Visitor: V (class_person)
Has−location: P (class_place)
Start−time: ST (class time_point)
End−time: ET (class time_point)
Has−duration: D (class duration)
Figure 7: Concept node for the visiting event
cording with the following criteria: how many feature for each type
were encountered in the story.
Badger obtained a case frame instantiations for Place and Vis-
itor using conceptual nodes defined in the dictionary constructed Figure 8: Badger output
by Crystal. In the Badger’s output the following conventions were
used: the name of the slot appears in the left hand side of the arrow
ID: 80 Type: conferring-a-monetary-award
and the value for the slot on the right hand side of the arrow. In
docid = ibrow-story
David Brown story, Badger instantiated Place to The OU and visi-
sentence_num = 1
tor to David Brown. The type of event is obtained from the value
segment_num = 1
of Type and the document ID from docid.
Funder ==> PP: FROM THE EUROPEAN COMMISSION
The output shown in Figure 8 means that Badger had instantiated
(using the CN definitions and domain lexicon) to a frame of the
form: ID: 106 Type: conferring-a-monetary-award
Concept Node: docid = ibrow-story
CN-type: visiting-a-place-or-people sentence_num = 1
Slots: segment_num = 1
Visitor tag: VI Money ==> OBJ1: 1 MILLION ECU
Start-time tag: ST
End-Time tag: ET
ID: 24 Type: conferring-a-monetary-award
Place tag: PL
docid = ibrow-story
Research-group tag: GR
sentence_num = 1
Date is not stated in the story. So Start-time and End-time are segment_num = 1
instantiated to the date in which the story was written. Project-Institution ==> SUBJ: IBROW
6. INFERENCE CAPABILITIES BY USING In this last example, we need to use the KMi planet ontology to
AN ONTOLOGY find if Project-Institution is a institution name or a project name,
An example of an story belonging to the type of event conferring- and this is done by a simple traversal of the inheritance links in
a-monetary-award is defined as follows. This example is described the ontology. Specifically, to remove ambiguity we sent a query to
in this paper because shows the inference capabilities which could Web-onto asking for the set of all educational-organizations using
be obtained from using an IE component plus an ontology. the following query code.
IBROW has been awarded 1 million Ecu from the Eu- web-onto display akt-kmi-planet-kb
ropean Commission to carry out research in the area of ocml-eval(setofall ?x
knowledge-based systems. (educational-organization ?x))
The output from Badger is shown as below. This gives a list containing all educational-organizations:
to give @(the-open-university 7. OCML CODE GENERATED FROM OUR
... SYSTEM
org-knowledge-media-institute)
Our goal is to use the information obtained by Badger and KMi
ontology in order to be able to populate our KMi ontology with
IBROW does not match any of these, however, we also send a
new instances of classes. In order to accomplish this task we had
query to Web-onto asking for the set of all kmi-projects:
plugged another component which is a translator from Badger’s
web-onto display akt-kmi-planet-kb output to OCML code. The main function of this translator is to
ocml-eval(setofall ?x tokenise the Badger output and then find the CN definitions (cn
(kmi-project ?x)) markers) and extract all the objects encountered in the story. The
name of each slot in the frame case corresponds to the name of the
yielding field in the class definition and the value for the field is the extracted
information.
to give @(project-d3e For the example David Brown’s story we end up with a visiting-
... a-place-or-people event and produce the intermediate output:
project-kmi-planet
... (def-instance visit-of-david-brown-
project-ibrow the-chairman-of-the-university
... visiting-a-place-people
project-heronsgate-mars-buggy) ((has-duration 1-day)
(start-time wed-15-oct-1997)
(end-time wed-15-oct-1997)
and hence a match of “IBROW” to project-ibrow
(has-location the-ou )
In a similar fashion a query is sent to webonto in order to find if
(visitor david-brown-the-
Funder is a valid funder body.
chairman-of-the-university)
)
web-onto display akt-kmi-planet-kb
)
ocml-eval(setofall ?x
(awarding-body ?x)) where an instance of the type event visiting-a-place-or-people
has been defined with the name “visit-of-david-brown-the-chairman-
to give @( ... of-the-university”.
org-european-commission
org-british-council)
8. POPULATING THE ONTOLOGY
At same time some semantic relations could be obtained by Building domain-specific ontologies often requires time-consuming
using the KMi planet ontology. For our example about IBROW we expensive manual construction. Therefore we envisage IE as a
can derive the following semantic relations: technology that might help us during ontology maintenance pro-
“ibrow is KMi project” and “KMi is part-of the Open-University” cess. During the population step our IE system has to fill prede-
The OCML query to derive that KMi is part of the open univer- fined slots associated with each event, as already defined the on-
sity is as follows: tology. Our goal is to automatically fill as many slots as possible.
However, some of the slots will probably still require manual inter-
web-onto display akt-kmi-planet-kb vention. There are several reasons for this problem:
ocml-eval(setofall ?x
there is information that is not stated in the story,
(organization-unit-part-of ?x
none of our templates match with the sentence that might
the-open-university)) provide the information (incomplete set of templates)
We note that there are some cases when the instances are not
to give @(knowledge-media-institute defined in the ontology and then determining the type of an object is
acad-unit-department-of-earth-science not straightforward. This has to be derived from a proof. Currently,
acad-unit-department-of-statistics-ou we still looking to this aspect of our research.
acad-unit-faculty-of-maths-and-computing-ou Figure 9 shows the extracted information from David Brown
... story.
org-office-for-technology-development) Once the system had extracted the information the user will pre-
sented with all extracted information even the one that cannot be
therefore we could conclude that: categorized as belonging to a type of object defined in our domain.
“the Open-University has been awarded 1 million Ecu from the Therefore, before populating the ontology we will require that a
European Commission” person check/complete the extracted information.
In a future implementation we will be interested in finding more
complex relations by using our KMi Planet ontology. 9. CONCLUSIONS AND FUTURE DIREC-
Finally, we remark that OCML (the query language used by we-
bonto) has adopted the closed world assumption (CWA), in the TIONS
same fashion as Prolog, and so facts that are not provable are re- We had built a tool which extracts knowledge using an ontology,
garded as “false” as opposed to “unknown”. an IE component and OCML translator. Currently, our system had
Figure 9: Extracted information
been trained using the archive of 300 stories that we had collected [2] M. Craven and J. Kumlien. Constructing Biological Knowledge Bases by
in KMi. 2 The training step was performed using typical examples Extracting Information from Text Sources. In Proceedings of The 7th
International Conference on Intelligent Systems for Molecular Biology
of stories belonging to each of the different type events defined in (ISMB-99), 1999.
the ontology. We obtained results over 95% using the IE compo- [3] J.B. Domingue. Tadzebao and WebOnto:Discussing, Browsing and Editing
nent in KMi stories. However, in the future we would like to use Ontologies on the Web. In Proceedings of the Knowledge Acquisition
the IE component in a different domain. We are interested in using Workshop, 1998.
[4] J-U. Kietz, A. Maedche, and R. Volz. A method for semi-automatic ontology
our system in companies project reports, Curriculum Vitae (CV’s), acquisition from a corporate intranet. In Proceedings of the EKAW’00 Workshop
or application of jobs. on Ontologies and Text, Juan-Les-Pins, France, oct 2000.
Another possible direction that we would like to explore is to [5] S. Luke, L. Spector, D. Rager, and J. Hendler. Ontology-based Web agents. In
incorporate into the IE component a different Machine Learning Proceedings of the First International Conference on Autonomous Agents,
pages 59–66. ACM, 1997.
algorithm such as described in [12]. in order to compare perfor- [6] A. Maedche and S. Staab. Semi-automatic engineering of ontologies from texts.
mance between them. In Proceedings of the 12th International Conference on Software Engineering
As medium term goal, we would like to have access to a library and Knowledge Engineering, SEKE2000, Chicago, IL, USA, pages 231–239, jul
2000.
of IE methods and to activate these over a web page or a collection [7] T. Mitchell. Generalization as search. Artificial Intelligence, 18:203–226, 1982.
of web pages. [8] E. Motta. Reusable Components for Knowledge Modelling. IOS Press,
Besides the above issues, Badger could be extended in order to Netherlands, 1999.
save its output in XML (Extensible Markup Language). This will [9] D. Proux and Y. Chenevoy. Natural Language Processing for Book Storage:
increase the portability of our IE system as XML is the universal Automatic Extraction of Information from Bibliographic Notices. In
Proceedings of The Natural Language Processing Pacific Rim Symposium
format for structured documents and data on the Web. (NLPRS’97), pages 229–234, 1997.
Finally, we would like to integrate our IE component with vi- [10] E. Riloff. An Empirical Study of Automated Dictionary Construction for
sualisation component. This visualisation component will allow Information Extraction in Three Domains. AI Journal, 85:101–134, 1996.
visualisation of all entities extracted. [11] C. Roux, D. Proux, F. Rechenmann, and L. Julliard. An Ontology Enrichment
Method for a Pragmatic Information Extraction System gathering Data on
Genetic Interactions. In Proceedings of The 14th European Conference on
10. ACKNOWLEDGMENTS Artificial Intelligence (Workshop on Ontology Learning ECAI-2000), 2000.
[12] S. Soderland. Learning Information Extraction Rules for Semi-structured and
The research described in this paper is supported by (EPSRC) free Text. Machine Learning, 34:1–44, 1999.
under the project name: Advanced Knowledge Technologies (AKT). [13] S. Soderland, D. Aronow, D. Fisher, J. Aseltine, and W. Lehnert. Machine
Learning of Text Analysis Rules for Clinical Records. Tr 39, Center for
Intelligent Information Retrieval, 1995.
11. REFERENCES [14] S. Staab, A. Maedche, and S. Handschuh. An annotation framework for the
[1] M. Craven, D. DiPasquo, D. Freitag, A. McCallum, K. Nigam T. Mitchell, and semantic web. In Proceedings of the First Workshop on Multimedia Annotation,
S. Slattery. Learning to Construct Knowledge Bases from the World Wide Web. 2001.
Artificial Intelligence, 1999.
URL:http://kmi.open.ac.uk/planet/