=Paper=
{{Paper
|id=Vol-184/paper-6
|storemode=property
|title=Keyword Extraction from the Web for Personal Metadata Annotation
|pdfUrl=https://ceur-ws.org/Vol-184/semAnnot04-06.pdf
|volume=Vol-184
|dblpUrl=https://dblp.org/rec/conf/semweb/MoriMIF04
}}
==Keyword Extraction from the Web for Personal Metadata Annotation==
https://ceur-ws.org/Vol-184/semAnnot04-06.pdf
Keyword Extraction from the Web
for Personal Metadata Annotation
Junichiro Mori1,3 , Yutaka Matsuo2 , Mitsuru Ishizuka1 , and Boi Faltings3
1
University of Tokyo, Japan
jmori,ishizuka@miv.t.u-tokyo.ac.jp
2
National Institute of Advanced Industrial Science and Technology, Japan
y.matsuo@carc.aist.go.jp
3
École Polythechnique Fédérale de Lausanne, Switzerland
junichiro.mori, boi.faltings@epfl.ch
Abstract. With the currently growing interest in the Semantic Web and Social
Networking, personal metadata is coming to play an important role in the Web.
This paper proposes a novel keyword extraction method to extract personal meta-
data from the Web. The proposed method is based on co-occurrence information
of words. Our method extracts relevant keywords depending on the context of
a person. Our experimental results show that extracted keywords are useful for
personal metadata creation. We also discuss the annotation of personal metadata
and application to the Semantic Web.
1 Introduction
The Semantic Web[2] is a new paradigm which brings “structure” to the meaningful
content of the Web. With currently growing interest in the Semantic Web and new stan-
dards for metadata description such as the Resource Description Framework (RDF)[13],
metadata is gradually gaining popularity in the Web.
Another recent trend in Web development is “Social Networking”[7]. Social Net-
working sites are community sites through which users can maintain an online network
of friends or associates for social or business purposes. Numerous Social Networking
sites have been launched recently.
As seen in Social Networking, a user itself is gradually coming to play a central role
in the Web contents (e.g. In “Weblog”, variety of contents is created by a user). With
these recent Web trends, expressing metadata about people and the relations among
them is recently gaining interest. In fact, some vocabularies and frameworks for per-
sonal metadata description have being developed [5][9][15][16].
Using these vocabularies, a user is gradually creating his or her personal metadata.
However, as a major problem of the Semantic Web is the metadata annotation, personal
metadata must also overcome the problem and need methods that facilitate and accel-
erate metadata annotation [8][10]. Although there are some supporting tools to create
personal metadata such as Foaf-a-Matic4 , this tool facilitates only basic descriptions.
4
http://www.ldodds.com/foaf/foaf-a-matic.html
51
Considering personal metadata, we notice that a lot of information is contained in
the Web pages. For example, imagine a researcher: that researcher’s information can
be in an affiliation page, a conference page, an online paper, or even in a Weblog. In
fact, we can expect that these pages contain a lot of personal metadata even including
information that we would not expect to find. Therein, questions are:
– What kind of personal metadata are in the Web?
– What kind of Web page contains personal metadata?
– How are extracted metadata applied to semantic annotation?
Considering these points, one of our research goals is to extract personal metadata
from the Web and apply them to semantic annotation. As a preliminary report to achieve
this goal, we propose a novel keyword extraction method to extract personal information
from the Web.
The remainder of this paper is organized as follows: section 2 describes the pro-
posed keyword extraction method using an actual example. In section 3, we show the
extracted keywords and analyze them. In section 4, we discuss the annotation of per-
sonal metadata. Section 5 contains related works. Finally, we address future works and
conclude this paper in section 7.
2 Keyword Extraction
2.1 Extraction of the Initial Term Set for Keyword
As an experimental attempt, we extracted the keywords of Program Committee mem-
bers of SemAnnot 2004 Workshop (There are 28 members including chair persons).
First, we need to acquire Web pages that contain information of respective committee
members and their mutual relationships. A simple way of acquiring those Web pages is
to use a search engine. It is reasonable to use a search engine because it can search many
Web pages in less than a few seconds. It also tracks the temporal variance of the Web. In
this experiment, we used Google5 , which currently addresses data from 4 billion Web
pages.
We first put each person’s full name to a search engine (name is quoted with dou-
ble quotation such as “Siegfried Handschuh”) and retrieve documents related to each
person. From the search result, we used the top 10 documents per person as the initial
documents that might contain personal keywords.
The search result documents include not only html files but also other file types
such as .pdf, .doc, .xls, .ppt. In this experiment, we used only html files. Furthermore,
we did not use metadata indicators in an html file such as META tags and RDF. In the
future, we are planning to use other file types along with html files that already have
been attached metadata.
The html files, at to a maximum of 10 files per person, are acquired from the initial
documents of each person. They are pre-processed with html-tag deletion and part-of-
speech tagging (POS). Then, the term set for keyword extraction is extracted from pre-
processed html files using the term extraction tool, Termex [14]. Termex extracts terms
5
http://www.google.com
52
from POS data based on statistical information of conjunctions between parts of speech.
Termex6 can also extract nominal phrases that include more than two nouns such as
“Annotation tool”. After the whole procedure of extracting the term set, we extracted
about 1000 terms per person on the average. The relevant keyword of each person is
chosen from these terms. Figure 1 shows steps of the proposed keyword extraction.
Fig. 1. Procedure of keyword extraction
2.2 Keyword Extraction Using Co-Occurrence Information
Because the term set includes both relevant and irrelevant terms for personal informa-
tion, we need to evaluate the relevance of term as a personal keyword. This subsection
explains the scoring method that gives relevance as a personal keyword to the term.
Term relevance based on Co-Occurrence The simple approach to measure term rel-
evance as a personal keyword is to use co-occurrence. In this paper, we define co-
occurrence of two terms as term appearance in the same Web page. If two terms co-
occur in many pages, we can say that those two have a strong relation and one term is
relevant for another term. This co-occurrence information is acquired by the number of
retrieved documents of a search engine result. For example, assume we are to measure
the relevance of name N (e.g. “Siegfried Handschuh”) and term w (e.g. “Annotation”).
Here, w is the term in the term set W extracted from the initial documents of the per-
son named “N ” . We first put a query, “N and w”, to a search engine and obtain the
number of retrieved documents that is denoted by |N and w|. We continuously apply a
query, “N ” and “w”, and obtain the number of retrieved documents for each, |N | and
|w|. Then, the relevance between the name N and the term w, denoted by r(N, w), is
6
Termex can be used for both Japanese and English POS data
53
approximated by the following Jaccard coefficient.
|N and w|
r(N, w) =
|N | + |w| − |N and w|
This Jaccard coefficient captures the degree of co-occurrence of two terms by their
mutual degree of overlap.
Keyword of person As described in a previous subsection, the term set of a person is
extracted from various Web pages. Although the Web pages contain a person’s name in
the text, each page may contain personal information in different contexts. For example,
imagine that one person, named “Tom”, is both a researcher and a artist, we can expect
that his name may appear not only in academic-related pages, but also in other pages
related to his art activities. Even among his academic-related pages, there might be
different pages depending on his acquaintances, affiliations, and projects. In this way,
different Web pages reflect different contexts of a person. Here, we introduce the notion
of a context to extract the keyword that captures the context of a person.
To extract the keyword in relation to a certain context, we must estimate the rele-
vance between the term and the context. If we replace the name N with the context C
in the relevance, r(N, w), we can obtain the relevance between context C and term w,
r(C, w), in the same manner. Then, the relevance of person N and term w in the context
C, denoted by score(N, C, w), is calculated as the following.
r(N, w) r(C, w)
score(N, C, w) = +α
M AX(r(N, w)) M AX(r(C, w))
r(N, w)
( > threshold)
M AX(r(N, w)
Therein, α denotes the relevance between the person and the context. For example, we
can use r(N, C) as α. M AX(r(X, Y )) is the maximum value of the Jaccard coefficient
in the term set W . We define the threshold for r(N, C) to exclude terms that are not
relevant for a person, but that have strong relation to the context. threshold is decided
based on heuristic method. The term w with the higher score(N, C, w) is considered to
be a more relevant keyword for person N in context C.
Regarding the “Tom” example, if we set “Art” as the context, we can get keywords
related to his art activities. Alternatively, if we include his research project name as the
context, keywords related to his project would be acquired.
Keywords showing a relation between persons If we consider the relation between
two persons in terms of their contexts, one person can be regarded as a part of the con-
text of another person. Hence, we can apply the previous formula to keyword extraction
of the relations among persons as follows:
r(N 1, w) r(N 2, w)
score(N 1, N 2, W ) = +β
M AX(r(N 1, w)) M AX(r(N 2, w))
r(N 1, w) r(N 2, w)
( , > threshold)
M AX(r(N 1, w) M AX(r(N 2, w)
54
Therein, N 1 and N 2 denote each person’s names in the relation. β is the parameter of
relevance between persons, such as r(N 1, N 2). This formula shows the relevance of
person N 1’s term w in relation to person N 2.
As there are many contexts of a person, the relations among persons also have a
variety of contexts. For example, the relation of two persons in the academic field
might be coauthors, have the same affiliation, the same project; they may even be
friends. The relevance of person N 1’s term w in relation to person N 2 in the context
C, score(N 1, N 2, C, w), is given as follows:
r(C, w)
score(N 1, N 2, C, w) = score(N 1, N 2, w) + γ
M AX(r(C, w))
Therein, γ is the parameter of relevance between the persons and the context, such as
r(N 1andN 2, C).
3 Keyword Analysis for Personal Metadata
3.1 Personal Metadata in Keywords
As an example of extracted keywords, Table. 1 shows higher-ranked extracted keywords
of “Siegfried Handschuh”. Each column in the table shows higher-ranked keywords
based on Term Frequency Inverse Document Frequency (TFIDF), co-occurrence with-
out the context, and co-occurrence with the context, respectively, from the left column.
In TFIDF-based keywords, we can find keywords that are related to the person such
as “annotation” and “semantic”. Nevertheless, there are many irrelevant words includ-
ing general words. Because TFIDF is based on the frequency of word appearances in a
text, it is difficult for a word to become higher-ranked in terms of relevance with another
word. On the other hand, in co-occurrence-based keywords, general words are excluded
and relevant words of each person appear in the rank list.
As explained in the previous section, the context can be considered in the keyword
extraction. In this experiment, we used “Semantic Web” as the context. With this con-
text, keywords are chosen in relation to one’s activity about the Semantic Web. In the
column of “Co-Occurrence with the context”, we can find that context-related keywords
come to appear in the rank list. The order of higher-ranked keywords also changes in
relation to the context.
The column at the right side shows a property label for each keyword in “Co-
Occurrence with the context”. Considering a correspondence to existing personal meta-
data vocabularies such as FOAF, we have defined six property labels: Name (N), Tech-
nical term (T), Event (E), Organization (O), Project (P), URL. In order to analyze what
kind of property is included in keywords, we annotated a property label to higher-ranked
keywords of each person. Thereby, we acquired 1646 labeled keywords in total (about
60 keywords per person on average).
Table. 2 shows the distribution of property labels. Nearly half of higher-ranked key-
words are occupied with names. Notwithstanding, it is noteworthy that other proper-
ties such as organizations and projects also appear to a certain degree. In particular,
as shown on the right side column, the properties for each person are distributed in a
balanced manner. This distribution indicates that if we extract about 60 higher-ranked
55
Table 1. Higher-ranked keywords of “Siegfried Handschuh” using TFIDF and co-occurance-
based method
TFIDF Co-Occurrence Co-Occurrence Property
(without the context) (with the context “Semantic Web”)
Semantic Siegfried Handschuh Siegfried Handschuh N
Siegfried Handschuh Ljiljana Stojanovic Ljiljana Stojanovic N
Office Nenad Stojanovic Nenad Stojanovic N
annotation Marc Ehrig Steffen Staab N
Person Julien Tane Marc Ehrig N
Web Steffen Staab Julien Tane N
Karlsruhe Daniel Oberle Daniel Oberle N
Konstanz Valentin Zacharias Valentin Zacharias N
E223 Andreas Hotho Andreas Hotho N
CREAM relational metadata Semantic Web T
karlsruhe.de annotation of web pages relational metadata T
message Knowledge Markup annotation of web pages T
Inf.wiss Large Scale Semantic Web Knowledge Markup T
knowledge automatic CREAtion of Metadata Large Scale Semantic Web T
Webmaster Annotation Workshop Knowledge Markup Workshop E
Appointment Knowledge Markup Workshop International Semantic Web Conference E
AIFB KCAP KCAP E
Katarina Stanoevska AIFB AIFB O
Beat Schmid University of Karlsruhe University of Karlsruhe O
Alexander Maedche OntoAgents OntoAgents P
keywords of one person, on average we can obtain about 30 names of his acquaintance,
2 or 3 related organizations, and 1 or 2 projects. These numbers nearly match our re-
search activity and show the possibility of using keywords for personal metadata. In
this analysis, we took many keywords together as “technical terms”. If we classify each
keyword more precisely, we could discover other personal metadata in keywords.
3.2 Personal Metadata in the Web
To further explore the possibility of personal metadata extraction from the Web, we
analyzed which Web pages include a higher-ranked keyword. First, we classified all
280 Web pages (10 per person) that were used to extract the initial term set. Thereby,
we prepared the 11 categories shown in Table. 3. “Personal page” includes personal
Web pages of the affiliation or one’s own domain. “Other page” includes uncategorized
pages and non-html pages such as .pdf and .ppt files. “Event page” includes conference,
workshop, and meeting pages. ML log is the email exchanged in a mailing list. DBLP 7
is the online bibliography of Computer Science papers. As seen in the table, “Personal
page” is the most dominant type of Web page. Because a person’s name was used as a
query, it is natural that we obtain a personal page in a search result.
7
http://www.informatik.uni-trier.de/l̃ey/db/
56
Table 3. Classification of the Web page type
Table 2. Distribution of properties labeled to
Web Page Number
higher-ranked keywords
Personal page 73 (26.0%)
Property Number Per person Other page 42 (15.0%)
Event 32 (11.4%)
Name 767 (46.5%) 27.3
ML log 27 (9.6%)
Technical term 613 (37.2%) 21.8
Online paper 26 (9.2%)
Event 105 (6.3%) 3.7
DBLP 22 (7.8%)
Organization 73 (4.3%) 2.6
Organization 17 (6.0%)
Project 48 (2.5%) 1.7
Project 16 (5.7%)
URL 40 (2.4%) 1.4
Book 11 (3.9%)
Total 1646
Publication list 8 (2.8%)
Weblog 6 (2.1%)
Total 280
Table 4. Distribution of each keyword property to each Web page type
Web page Name Technical Event Organization Project URL
Term
Personal page 234 (19.3%) 199 (24.0%) 31 (24.0%) 35 (36.8%) 30 (44.1%) 14 (25.9%)
Other page 42 (3.4%) 15 (1.8%) 4 (3.1.%) 3 (3.1%) 1 (1.4.%) 2 (3.7%)
Event 223 (18.3%) 171 (20.6%) 29 (22.4%) 25 (26.3%) 1 (1.4%) 14 (25.9%)
ML log 165 (13.6%) 122 (14.7%) 11 (8.5%) 16 (16.8%) 8 (11.7%) 8 (14.8%)
Online paper 12 (0.9%) 33 (3.9%) 4 (3.1%) 1 (1.0%) 1 (1.4%) 2 (3.7%)
DBLP 314 (25.9%) 189 (22.8%) 38 (29.4%) 0 11 (16.1%) 0
Organization 66 (5.4%) 45 (5.4%) 4 (3.1%) 8 (8.4%) 5 (7.3%) 9 (16.6%)
Project 46 (3.7%) 13 (1.5%) 0 5 (5.2%) 5 (7.3%) 5 (9.2%)
Book 18 (1.4%) 7 (0.8%) 1 (0.7%) 0 0 0
Publication list 85 (7.0.%) 24 (2.8%) 6 (4.6%) 1 (1.0%) 1 (1.4%) 0
Weblog 7 (0.5%) 10 (1.2%) 1 (0.7%) 1 (1.0%) 5 (7.3%) 0
Total 1212 828 129 95 68 54
Table 4 shows which category of Web page a higher-ranked keyword belongs in (a
keyword may appear in more than one category). Specifically examining each column,
we find which kind of Web page each property is included in. Moving the focus to a
row in the table, we can find what kind of property each Web page category includes.
Although name entities can be acquired most from “Personal page”, DBLP is also
a good information resource to extract a name entity. DBLP contains coauthor infor-
mation of a paper. Therefore, the extracted name is related to one’s acquaintance in a
research activity. “Event page”, such as conference, workshop, is a information resource
of various personal information. However, because the Event page is not specified to a
certain person, “Personal page” gives more accurate information about each person.
Overall, “Personal page” is a good information source for personal metadata such
as names, organizations, and projects. Event page and DBLP provide metadata that are
57
related to personal research activities such as coauthors, projects, and events including
conferences and workshops.
4 Annotation of Personal Metadata
Our keyword extraction method can be applied to semantic annotation in following
ways.
– Annotation for Web page : As our analysis showed, our personal keyword extrac-
tion method offers strong potential for personal metadata extraction from the Web.
Extracted personal metadata can be applied to partially annotate the Web pages us-
ing metadata description framework such as the RDF[13]. Because metadata are
given the relevancy in relation to a person, annotated Web pages can be used in
many applications such as Information retrieval and Information integration. For
example, using annotated Web pages, the search engine that supports the Semantic
Web could answer to following question:
• Who knows this person?
• Who is involved in this project?
• Who knows this research topic well?
• Which pages include this person’s information?
– Annotation for Personal Metadata File : Extracted personal metadata is used not
only for annotating a Web page, but also for annotating a personal metadata file. As
one emerging personal metadata standard, “Friend of a Friend”, FOAF[5], defines
an RDF vocabulary for expressing metadata about people, the relation among them,
and the things they create and do. FOAF provides a way to create machine-readable
personal documents on the Web, and to process them easily through merging and
aggregating them. Because extracted metadata are easily incorporated in FOAF, we
can facilitate the creation of FOAF documents.
This paper presents discussion of the importance of a person’s context in keyword ex-
traction. The context often defines the properties. Currently, there is no FOAF vocabu-
lary to define a context. In addition to FOAF, there are many vocabularies and frame-
work for personal metadata such as Topicmaps [9], RDF-vCard [16], Person class of
DAML+OIL [15]. However, none of them address the notion of a personal context.
One way to introduce a personal context to those metadata frameworks is to prepare
schema that corresponds to respective contexts. Regarding the expression of personal
metadata, we need further consideration to make the metadata expressive and usable.
5 Related works
Aiming at extracting and annotating personal metadta, our method is regarded as one
of Information Extraction(IE) methods supporting a semantic annotation. Up to now,
many IE methods rely on predefined templates and linguistic rules or machine learning
techniques to identify certain entities in text documents[12]. Furthermore, they usually
define properties, domains, or ontology beforehand. However, because we try to extract
various information from different Web pages, we don’t use predefined restrictions in
the extraction.
58
Some previous IE researches have addressed the extraction and annotation of per-
sonal metadata. In [1], they proposes the method to extract a artist information, such
as name and date of birth, from documents and automatically generate his or her biog-
raphy. They attempt to identify entity relationships, metadata triples (subject-relation-
object), using ontology-relation declarations and lexical information. However, Web
pages often include free texts and unstructured data. Thereby, capturing entity relation-
ships becomes infeasible because of lacking regular sentences. Rather than focusing on
the entity relationship, we find the entity in the Web pages based on the relevance in
relation to a person.
In [6], they address the extraction of personal information such as name, project,
publication in a specific department using unsupervised information extraction. It learns
to automatically annotate domain-specific information from large repositories such as
the Web with minimum user intervention. Although they extract various personal meta-
data, they don’t consider the relevance of extracted metadata. Because extracted meta-
data in our method have the relevance, they can be used as reliable initial seeds for
bootstrap learning for automatic annotation in their method.
Although the aim is not extracting personal metadata, in [11], they proposes the
method to extract a domain terminology from available documents such as the Web
pages. This method is similar to our one in terms of that terminology are extracted
based on the scoring measure. However, their measure is based not on the co-occurrence
but on the frequency. Furthermore, they focus on the domain-specific terms rather than
personal metadata and the method is domain dependent. In our method, we can capture
the various aspects of personal metadata even from different domain resources using
the notion of a context.
6 Future works and Conclusion
To apply our keyword extraction methods to personal metadata annotation, we must
consider and solve following points in the future.
– Evaluation of personal metadata : One problem is that we are not sure that the
extracted metadata are true. Although two terms co-occur in many Web pages, they
might not have any relation. Therefore, someone should evaluate the propriety of
a keyword as actual metadata. One approach to solve this problem would be an
interactive annotation system[3]. Reusing and modifying a keyword as a candidate
of personal metadata, a user can easily annotate personal metadata.
– Entity recognition of keywords : Another critical problem is to decide a certain
keyword property. In our experiment, the property label was given manually to
each keyword. However, it is not efficient to put a property to numerous extracted
keywords. One approach to automatically decide the property of a keyword is to
use techniques in the entity recognition research[4].
– Privacy problem of information extraction from the Web : A person sometimes
does not know that his or her information is extracted from the Web only by name.
Therefore, we should take care not to intrude on a user’s privacy even in information
extracted from the Web. We must clarify the use of the information only for useful
services for a user.
59
The Web holds much personal information that can be used as personal metadata.
This paper proposes a novel keyword extraction method to extract personal information
from the Web. Our result showed the important possibility of using extracted keywords
as personal metadata. Importantly, our method can capture the personal information in
different contexts. This allows us to obtain various personal metadata.
Because the Web is such a large information resource, its information runs the
gamut from useful to trivial. It presents the limitation that it must be publicly avail-
able on the Web. For further improvement of the proposed method, we must analyze
“what” information of “who” in the Web, and its reliability.
References
1. H. Alani et al. Automatic Extraction of Knowledge from Web Documents. In Workshop of
Human Language Technology for the Semantic Web and Web Services, 2nd International
Semantic Web Conference, Sanibel Island, Florida, USA, 2003.
2. T. Berners-Lee, J. Hender, O. Lassila. The Semantic Web. Scientic American, 2001.
3. F. Ciravegna, A. Dingli, D. Petrelli, and Y. Wilks. User-system cooperation in document
annotation based on information extraction. In Proceedings of the 13th International Con-
ference on Knowledge Engineering and Knowledge Management, EKAW02, Springer Verlag,
2002.
4. H. Cunningham et al. GATE:A Framework and Graphical Development Environment for
Robust NLP Tools and Application. In Proceedings of the 40th Anniversary Meeting Assoc.
for Computational Linguistics(ACL2002), East Stroudsburg, Pa., 2002.
5. Dan Brickley and Libby Miller. FOAF: the ’friend of a friend’ vocabulary. http://xmlns.
com/foaf/0.1/, 2004.
6. A. Dingli, F. Ciravegna, D. Guthrie, Y. Wilks. Mining Web Sites Using Unsupervised Adap-
tive Information Extraction. In Proceedings of the 10th Conference of the European Chapter
of the Association for Computational Linguistics, Budapest, Hungary, 2003.
7. L. Garton, C. Haythornthwaite, and B. Wellman. Studying online social networks. In Doing
Internet Research, S. Jones, Ed. Sage, Thousand Oaks, CA, pp. 75–105, 1999.
8. J. Kahan and M. R. Koivunen. Annotea: An open rdf infrastructure for shared web annota-
tions. In Proceedings of the 10th International WWW Conference, pp.623–632, 2001.
9. Lars Marius Garshol. Living with topic maps and RDF.
http://www.ontopia.net/topicmaps/materials/tmrdf.html, 2003.
10. S. Staab, A. Maedche, and S. Handschuh. An Annotation Framework for the Semantic Web.
In Proceedings of 1st International Workshop MultiMedia Annotation, 2001.
11. P. Velardi, M. Missikoff, R. Basili. Identification of relevant terms to support the construc-
tion of Domain Ontologies. In ACL-EACL Workshop on Human Language Technologies,
Toulouse, France, 2001.
12. R. Yangarber and R. Grishman. Machine Learning of Extraction Patterns from Unanotated
Corpora: Position Statement. Workshop Machine Learning for Information Extraction, IOS
Press, Amsterdam, pp.76–83, 2000.
13. Resource Description Framework(RDF) Schema Specification. In W3C Recommendation,
2000.
14. http://gensen.dl.itc.u-tokyo.ac.jp/win.html
15. DAML Ontology Library. http://www.daml.org/ontologies/
16. Representing vCard Objects in RDF/XML. http://www.w3.org/TR/2001/NOTE-vcard-rdf-
20010222/
60