Reck, Ronald P., Kenneth B. Sall and Wendy A. Swanbeck. “Determining the Impact of Eric Clapton on Music Using RDF Graphs: Selected Challenges
of Semantics Across and Within Datasets.” Presented at Balisage: The Markup Conference 2011, Montréal, Canada, August 2 - 5, 2011. In Proceedings of Balisage: The Markup Conference 2011. Balisage Series on Markup Technologies, vol. 7 (2011). https://doi.org/10.4242/BalisageVol7.Sall01.
Balisage: The Markup Conference 2011 August 2 - 5, 2011
Balisage Paper: Determining the Impact of Eric Clapton on Music Using RDF Graphs
Selected Challenges of Semantics Across and Within Datasets
For over a decade Ronald P. Reck has operated the consulting company, RRecktek LLC, outside of
Washington DC metropolitan area. RRecktek LLC has enjoyed over one hundred
contracts ranging from the data warehousing of state, local, and federal law
enforcement incident reports outside of submarine bases for The Navy Criminal
Investigative Service (NCIS) to vocabulary projects for the management and
dissemination of controlled vocabularies for the Directorate of National
Intelligence (DNI) as a member of the Intelligence Community Metadata Working
Group staff and a "simple" content management system for build out drawings for
global telecom company. Among the companies served include Nextel, Winstar, ANS
+COre, AOL, Standard & Poors, The Federal Communications Commission,
Kiplingers Newletter, Radio Free Asia, Eastman Kodak, The United States
Information Agency, The Council of Better Business Bureaus, Department of
Defense Health Affairs and others. He prides himself on developing scalable,
open source architectural strategies for difficult problems. Ron resides with his
lovely wife Olga and the best son in the entire world.
Ken Sall has been supporting the US federal government in XML efforts since
1998. His customers include NASA, General Services Administration (GSA),
Department of Homeland Security (DHS), and the Intelligence Community. Sall has been
an active contributor to XML and data standardization efforts including the Federal
Enterprise Architecture - Data Reference Model (FEA DRM), the National Information
Exchange Model (NIEM), and the Intelligence Community Metadata Standards for
Publication (IC MSP), as well as participating in several federal XML and data
management working groups. As the author of
XML Family of Specifications: A
Practical Guide (Addison-Wesley, 2002), he basks in the glory of quarterly
statements from his publisher that no longer include payments. Music is his
passion. XML too.
Wendy Swanbeck has worked as a software engineer for over 20 years. In the
past she has worked on a variety of projects including graphical design
programs, mainframe control systems, and CAD design software for commercial and
government projects. More recently she has been working at Eastman Kodak Company
writing software for networking systems, color management and photo manipulation
GUI software. She also donates some of her time creating websites for groups
that need it. Her passion is to architect and write clean, flexible and robust
software using the right tools for the job.
As music is a topic of interest to many, it is no surprise that
developers have applied web and semantic technology to provide various RDF datasets
for describing relationships among musical artists, albums,
songs, genres, and more. As avid fans of blues and rock music, we
wondered if we could construct SPARQL queries to examine properties and
relationships between performers in order to answer global questions
such as "Who has had the greatest impact on rock music?" Our primary
focus was Eric Clapton, a musical artist with a decades-spanning career
who has enjoyed both a very successful solo career as well as having
performed in several world-renowned bands.
The application of semantic technology to a public
dataset can provide useful insights into how similar approaches can be
applied to realistic domain problems, such as finding relationships
between persons of interest. Clearly understood semantics of available
RDF properties in the dataset is of course crucial but is a substantial
challenge especially when leveraging information from similar yet
different data sources.
This paper explores the use of DBpedia and MusicBrainz data sources
using OpenLink Virtuoso Universal Server with a Drupal frontend. Much
attention is given to the challenges we encountered, especially with
respect to relatively large datasets of community-entered open data sources of varying
quality
and the strategies we employed or recommend to overcome the challenges.
A long, long time ago in a galaxy much like our own, one of the authors, Ken Sall,
was fascinated by the number of musicians who had played with John Mayall. [1]
Armed only with paper, pencil and lots of record album jackets, he drew a diagram
with Mayall at the center and arcs pointing to musicians who played with him, such
as
Eric Clapton, Peter Green, Mick Taylor, etc. Then he drew arcs from these musicians
to others they
had played with. Soon he reached the limit of the paper. Since this was prior to the
emergence of the PC, there was no option to scroll or easily copy to a larger drawing
area. But Ken held onto this labor of trivia for nearly 40 years.
(See Figure 1.)
Meanwhile, Pete Frame developed far more detailed and visually appealing
rock family trees Frame 1983,
some of which appeared in various publications such as Rolling Stone magazine
and the Encyclopedia of Rock.
Nearly everyone is familiar with the Six Degrees of Kevin Bacon
trivia game Bacon.
The Bacon number is based on a game where actors were rated on distance they were
away from the actor
Kevin Bacon. Actors who appeared in a movie with Kevin Bacon were given a Bacon number
of 1, and actors
having only been in movies with actors with a Bacon number of 1 were given a Bacon
number of 2.
Each degree of separation lead to increasing the Bacon number by 1.
The object is to connect any movie star to Mr. Bacon with no more than six hops based
on
two people appearing in either the same movie or commercial.
This is a popular special case
of the more general Six Degrees of Separation Six Degrees
also known as the Human Web. The idea is that any two people on the planet can be
connected by no more than six hops.
According to Wikipedia," Mathematicians use an analogous notion of collaboration distance".
Figure 1: John Mayall Connections
Dozens of musicians are within a few degrees of separation from John Mayall (circa
1972).
Problem Statement
Using semantic technology, we can certainly improve upon hand copying of data from
record album jackets. If we can refer to a recording artist with all the associations
to
other musicians and to the albums they appeared on together, we can produce more
complete graphs than what is shown in Figure 1.
As avid fans of blues and rock music, we wondered if we could construct SPARQL queries
to examine properties and relationships between performers in order to answer global
questions such as "Who has had the greatest impact on rock music?" Our primary focus
was
Eric Clapton, a musical artist with a decades-spanning career who has enjoyed both
a
very successful solo career as well as having performed in several world-renowned
bands.
Then using Drupal and SVG to visualize the results, we could traverse the musician
graphs in a straightforward manner.
This paper explores the use of DBpedia and MusicBrainz data sources using OpenLink
Virtuoso Universal Server with a Drupal frontend. Much attention is given to the
challenges we encountered, especially with respect to community-entered open data
sources and the strategies we employed to overcome the challenges.
One such challenge we encountered was that there were several properties by which
an artist
could be connected to another and the semantics were not well-defined, as discussed
later.
We should be able to draw inferences from the data. According to Dean Allemang and
Jim Hendler:
In the context of the Semantic Web, inferencing simply means that given some stated
information,
we can determine other related information that we can also consider as if it had
been stated.
Which musician-related properties are reversible (inverse makes sense)?
How can we differentiate between a musician's playing in a band, being associated
with other musicians,
starring together in a live show, and others collaborating with the musician?
Does total number of songs or album released correlate with other measures of success?
Who created the most songs?
Which song has been recorded the most times by any artists?
("Yesterday" and "White Christmas" are typically cited.)
Is there a predominant record label in the music world?
Which solo artist has had the longest career?
Which band has been together (in some form) the longest time?
What is the average age of a musician when he/she first joined a band?
For bands with changing membership, can we conclude which configuration lasted the
longest?
What is the "Eric Clapton number" (a la Kevin Bacon number) for various musicians?
Can we use our own knowledge of Eric Clapton to clarify some of the semantics behind
some of the RDF data we encounter?
Our notion of a musical artist's activity and impact can be explained in general terms.
We consider
activity to be correlated with the number of recordings produced.
Another factor that we chose not to consider is an artist's concert performances
unless they resulted in a commercially available recording.
An artist's impact is more subjective. The greater the number of
musicians that play with a performer (the greater the number of associations), the
greater the
potential impact of the performer, provided that the performer is not simply a session
man.
Another measure that would have proved extremely helpful in determining both
activity and impact in a more quantitative manner would be the use of Billboard chart
data.
Unfortunately, use of Billboard data is not royalty free.[2]
Furthermore, our working definitions of activity and impact is primarily based on
commercial recordings of an artist.
We acknowledge that some artists tend to perform numerous live concerts and yet have
produced
relatively few commerical recordings; many of these concerts may be available on bootleg
recordings.
The argument could easily be made that the number of concerts and/or number of bootleg
recordings
are correlated with activity and/or impact. We have chosen not to consider these factors
at this time.
Data Sources
Fortunately in 2011, there is a tremendous wealth of information about musicians freely
available on the Web either in structured markup, especially triples, that lends itself
to
SPARQL queries. For example, Wikipedia tells us that Eric Clapton Clapton is associated with:
Dire Straits, The Yardbirds, John Mayall & the Bluesbreakers, Powerhouse, Cream, Free
Creek,
The Dirty Mac, Blind Faith, J.J. Cale, The Plastic Ono Band, Delaney, Bonnie & Friends,
Derek and the Dominos, and T.D.F.
By comparison, Ringo Starr Starr is associated with The Beatles, Rory Storm and the Hurricanes,
Ringo Starr & His All-Starr Band, and Plastic Ono Band. On the surface, Ringo's direct
associations are fewer,
but actually Ringo Starr & His All-Starr Band has had 11 different lineups (to date)
with a total of 42 unique musicians, most of whom have a number of associations as
well. Ringo Starr & His All-Starr Band
The various data sources we leveraged are discussed in the following subsections.
Wikipedia and Infoboxes
The primary (although indirect) source for our RDF data was Wikipedia which
is a major source of detailed information about musicians, among many other things.
To understand both
the kinds of properties available and their open community origins, some details about
Wikipedia are in order.
As illustrated by the excerpt from Clapton's Wikipedia page below,
the main page for a musical artist
contains an abstract at the top, a contents navigation box below the abstract with
a variable number
of section links pointing to the main content of the page,
and a so-called infobox in the upper right.
In addition to a main page,
most musical artists with more than a handful of albums or singles have a separate
discography page with
varying amount of detail and organization regarding studio albums, live albums, compilations,
singles, etc.
The discography is linked to the musician's main page and vice versa.
Clapton's Wikipedia page: Eric Clapton's Wikipedia Page (excerpt with Infobox on right)
Wikipedia. "Eric Clapton -- Wikipedia, The Free Encyclopedia". 2011.
An infobox is
a fixed-format table designed to be added to the top right-hand corner of Wikipedia
articles to consistently and concisely present a summary of some common aspects that
the articles of the same category (i.e., musical artist) share, as well as to
improve navigation to other interrelated articles (i.e., music genres). Infoboxes
are an instance of MediaWiki's template feature; there are numerous infobox
templates arranged by broad categories such as arts and culture infobox templates, which is further divided into 10
subcategories including templates for film, fictional characters, and music. There
are over 50 templates in the subcategory music
infobox templates of which the most relevant to our work is the template for the infobox of musicial artists. (See the right side of
Table 1 below.) This infobox template is used by
Wikipedia authors to create infoboxes such as Clapton's, shown in the left side of
the table. The
correspondence between the template and the resultant infobox is apparent when
viewed side by side. Most of the properties that we used in our queries (e.g., name,
genres, associated acts, etc.) are based on the contents of the infobox, with the
notable exception of albums (from the discography page).
Table 1
Clapton's Infobox (left) and Generic Musical Artist Infobox Template (right) [cited
03 Apr 2011]
The wiki source markup of the infobox is shown in Figure 3. For a much more complete
explanation of the Wikipedia extraction process employed by DBpedia including a discussion
of
the design and development of infobox templates,
see Auer and Lehmann 2007 and Auer et al 2007.
Figure 3: Wiki Source Markup for Clapton Infobox
Compare markup to rendering shown in Table 1.
DBpedia
One of the two primary dataset we used was DBpedia 3.6 en (English) based on Wikipedia
dumps from October/November 2010.
DBpedia DBpedia Dataset is a community effort to provide sophisticated query access to the structured content
of
Wikipedia, thereby allowing a small group of researchers and developers to enhance
Wikipedia by linking to additional datasets.
The DBpedia 3.6 release announcement
describes the content in detail:
"The new DBpedia dataset DBpedia Release describes more than 3.5 million things, of which
1.67 million are classified in a consistent ontology, including 364,000
persons, 462,000 places, 99,000 music albums, 54,000 films, 16,500 video
games, 148,000 organizations, 148,000 species and 5,200 diseases.
The DBpedia dataset features labels and abstracts for 3.5 million things in
up to 97 different languages; 1,850,000 links to images and 5,900,000 links
to external web pages; 6,500,000 external links into other RDF datasets, and
632,000 Wikipedia categories.
The dataset consists of 672 million pieces of information (RDF triples) out
of which 286 million were extracted from the English edition of Wikipedia
and 386 million were extracted from other language editions and links to
external datasets."
The DBpedia Ontology DBpedia Ontology is a cross-domain ontology
created from the most commonly used infoboxes within Wikipedia.
The ontology currently covers over 272 classes with 1,300 different properties and
1.667 million resources,
364,000 of which are Persons.
DBpedia predicate IRIs [3] begin with either http://dbpedia.org/ontology or
http://dbpedia.org/property; entitites are designated
by http://dbpedia.org/resource.
We initially used a simple SPARQL query SPARQL 1.0 to determine what properties are relevant to a musician
as shown in the following two listings. Some musicians had less than the 23 properties
shown;
others had more. Note that the resource IRI designating the musician is a straightforward rendering of
the artist's name in English with underscores replacing spaces:
http://dbpedia.org/resource/Eric_Clapton.
Figure 4: Display Properties Defined for Eric Clapton (as object)
SELECT DISTINCT (?predicate) WHERE {
?s ?predicate <http://dbpedia.org/resource/Eric_Clapton>.
} ORDER BY ?predicate
Clapton's DBpedia Properties - Version 1: Clapton's DBpedia Properties - Version 1 (23 Predicates)
In a later section, we discuss alternative property-related queries with different
results.
Wikipedia-like categories can also be specified in queries. The following query returns
the short list of musical artists who have three distinctions: Grammy Award winners,
Rock and Roll Hall of Fame inductees, and
MTV Video Music Awards winners.
SELECT ?allAwards
{
?allAwards <http://purl.org/dc/terms/subject>
<http://dbpedia.org/resource/Category:Grammy_Award_winners>.
?allAwards <http://purl.org/dc/terms/subject>
<http://dbpedia.org/resource/Category:Rock_and_Roll_Hall_of_Fame_inductees>.
?allAwards <http://purl.org/dc/terms/subject>
<http://dbpedia.org/resource/Category:MTV_Video_Music_Awards_winners>
} ORDER BY ?allAwards
The thrice-honored, distinguished musical artists are:
[4]
Figure 6: Dereferencing the Eric Clapton DBpedia IRI (excerpt)
MusicBrainz
MusicBrainz MusicBrainz is another comprehensive public database for musician information; it
contains detailed data about artists, release groups, releases, tracks, record labels
and many relationships between them.
As described in the database overview,
MusicBrainz defines artist attributes including the artist's name, aliases, GUID,
annotation, type (individual or group), and
begin and end dates (birth/death of an individual or formation/disbanding for a group).
A release group is a logical grouping
of variant releases (deluxe/limited editions, reissues/remasters, international variations,
box sets, etc.). One example is
the release group describing several
variants of Clapton's 461 Ocean Boulevard album.
Each release group is identified by type (album, single, EP, compilation, soundtrack,
live, etc.), as well as by artist, title, and ID.
An individual release has the same properties as a release group as well as status
(official, bootleg, etc.), Amazon ASIN,
annotation, language, release event (date, country, format, label, etc.) and more.
Each track has properties such as
duration (in milliseconds) and PUID (the MusicIP acoustic fingerprint identifier for
the track) as well as artist, title, and ID.
Record labels are also described with a number of detailed attributes.
Every MusicBrainz ID (MBID) is a permanent GUID. There is a direct relationship between
an artist's MBID, a IRI that
identifies the artist, and a web page that collects information about the artist.
For example:
Eric Clapton's MBID is 618b6900-0618-4f1e-b835-bccb17f84294
Below is a screenshot of the
Eric Clapton MusicBrainz page.
MusicBrainz offers a tremendous amount of detail about releases as exemplified by
the page for Clapton's classic
From the Cradle blues album.
Figure 8: Web Service Result - Query for 461 Ocean Boulevard
MusicBrainz maintains numerous database statistics,
a small sampling of which appear below. It is interesting to note that in the three
months since the draft version
of this paper, although the number of Recordings increased by approximately 37,000,
the Artist count dropped by roughly 8,000. This seems to suggest the dynamic nature
of the data both in terms
of quantity and in how recordings are categorized. (For example, artists may have
appeared under variant name
spellings.) We could not find a definition for the new term Works in the
terminology page or other
MusicBrainz documentation.
Table II
Selected MusicBrainz Statistics [accessed 10 April 2011; updated 26 June 2011]
Artists
612,428
Release Groups
787,918
Releases
952,743
Disc IDs
460,876
Recordings [tracks]
10,307,311
Labels
52,156
Works
276,864
Relationships [links]
3,160,096
The MusicBrainz data quality page
states that one of the goals is "Establish a method to determine the quality of an
artist and the releases that belong to that artist.
This provides consumers of MusicBrainz a clue about the relative quality rating of
the data in the database."
The page explains the connection between the quality number, voting periods, and strictness
regarding edits.
To accomplish these goals, this feature will allow editors to indicate the quality
for a given artist.
An artist can be of unknown, low, medium or high data quality. The data quality indicator
determines what
level of effort is required to change the artist information or to add/remove albums
from an artist.
An artist with unknown or medium quality will roughly require the amount of effort
that MusicBrainz currently
requires to edit the database. An artist with low data quality will make it easier
to add/remove albums or
to change the artist information (name, sortname, aliases). And an artist with high
data quality will
require more effort to add/remove albums or the change the artist information.
The data quality concept also applies to releases in the same manner. Changing a release
with low data
quality will be easier than changing a release with high data quality.
Freebase
Initially we considered using music data from Freebase Freebase, an open, community-based, Creative Commons licensed
repository of structured data describing millions of entities (i.e., person, place,
thing) which are connected as a graph. Freebase IDs can be used to uniquely identify
entities from any web-reachable source. At the time of this writing, the 1.22 GB
(uncompressed) music segment of Freebase contained 9 million topics and 33 million
facts. The music category contains classical, opera, and many other genres in
addition to blues and rock music. Data is formatted as approximately 50 separate Tab
Separated Values files with filenames such as group_member.tsv,
group_membership.tsv, musical_group.tsv, guitarist.tsv, album.tsv, artist.tsv,
release.tsv, and track.tsv (the largest file at roughly 671 MB). A total of over
151,000 groups are listed in musical_group.tsv. [5]
Freebase data can be used in several ways. MQL (Metaweb Query Language) MQL is a query API analogous to SPARQL for RDF. MQL uses JSON
objects as queries via standard HTTP requests and responses. For example, the IRI
below will return all the genres associated with Eric Clapton (line breaks added for
readability).
We will see later that this list of genres differs from other data
sources.
For any given group, band members are non-sequential within the
group_membership.tsv file. The relevant lines for the members of Blind Faith are
collected in the table below.
Table III
Members of Blind Faith - from
Freebase
id
member
group
(role)
start
end
/m/01tfhrr
Ginger Baker
Blind Faith
1968
1969-10
/m/01tfhry
Steve Winwood
Blind Faith
1968
1969-10
/m/01wvwr8
Ric Grech
Blind Faith
1968
1969-10
/m/01tfhrk
Eric Clapton
Blind Faith
1968
1969-10
The group_membership.tsv file contains dozens of entries for Clapton, one line
for each band or other association in which he participated. Each entry (line) is
identified by a different ID. For example:
/m/01t73cp Eric Clapton Derek and the Dominos
Eric Clapton's
main Freebase page is shown below. The IRI format for Clapton as a topic
is http://www.freebase.com/view/en/eric_clapton.
Figure 10: Eric Clapton's Freebase Page
Ultimately we elected not to use Freebase as a data source because we, like
others, were unable to locate the complete Freebase dataset rendered as RDF and
determined that the conversion process would be non-trivial. In fact, we found
mailing list and forum messages with others expressing the same problem. Initially
we had downloaded the dataset in its native format and considered converting it into
RDF. This would have been possible although many of the lines contained a fourth
element consisting of a number of values concatenated together; it was unclear how
that could be cleanly converted into a named graph. After the rough draft of our
paper was submitted, we discovered that RDF data is available manually by following
a link labeled "RDF" near the bttom of each Freebase page. It appears that
this RDF is rendered at execution time since there is a slight delay in the display
of the RDF. For example, follow the RDF link on the Eric
Clapton page. Had we discovered this sooner we might have attempted to
obtain the Freebase RDF for Rock and Roll stars using some automated process.
Methodology
In this section, we discuss our frontend and backend development platform, a few term
definitions,
and details concerning the Graphviz visualization which is still under development
as of this July writing.
Development Environment
The front end to our SPARQL queries was a guest virtual machine running under Sun
VirtualBox 3.1.6_OSE. on a Dell Inspiron with an Intel Core 2 Quad CPU Q9400 running
at 2.66GHz. The guest operating system is the 10.10 server release of Ubuntu Linux.
The
database machine leverages 2 Intel Xeon X5650 CPUs overclocked to 3.12GHz with 48
gigabytes of RAM. From the perspective of the 10.10 Ubuntu Desktop operating system,
the 2 hexcore processors are regarded as 24 processors (149867.81 BogoMIPS).
Our RDF store was the open source (freely-available) version of
OpenLink Virtuoso Universal Server Virtuoso Universal Server (Version 06.01.3127) running in "Full Mode". Configuration changes were made to
maximize use of available RAM by making a change to the virtuoso.ini file:
"NumberOfBuffers = 7000000".
On the frontend, our queries were facilitated by Drupal 6.18 with the modules listed
in the
Drupal Modules appendix.
It was necessary to configure PHP to allow Drupal more memory, so "memory_limit =
512M"
was added to /etc/php5/apache2/php.ini.
Drupal 7.0 became available during 2011, but we elected not to migrate to that version.
Although Drupal 7.0 does have substantially better support for RDF data than earlier
versions, that is likely to only effect people publishing information that was
authored or stored inside Drupal. Our use of Drupal was for collaborative development
of and storing of
queries, as well as for the visualization capabilities.
Working Definitions of Key Properties
The definition list below presents our current thinking about several key concepts
that
are covered in details in later sections; some refer to OpenCyc for the Semantic Web.
a person who is either a musician who plays one or more instruments,
or is a singer, or is a music composer; similar (but not identical) to
http://musicbrainz.org/mm/mm-2.1#Artist
OpenCyc: An element of Band_MusicGroup
is a (small or large) group of musicians who play
non-Classical music together on either a regular or intermittent schedule.
a categorization of music into types that can be distinguished from other types of
music;
can be applied to a musical artist, a band, an album, or individual songs.
RDF Visualization Coding
In order to better understand our result sets a form of visualization was
implemented as a custom Drupal module, written in PHP.
The front end was a simple HTML form with JQuery and AJAX for displaying the information.
Main processing was handled by the PHP module setup to intercept a post to a particular
URL.
The PHP module first accesses the POST data and extracts the input search term which
became the subject of the following SPARQL query:
SELECT DISTINCT ?predicate ?object
WHERE {<" . $subject . "> ?predicate ?object }
ORDER BY ?predicate
where in our initial implementation $subject is replaced by the musician subject requested by the user.
XML output is specified by the request as the desired return format.
The next step was to convert the XML results output from the query to the
Graphviz input format DOT language.
Graphviz Graphviz is open source graph visualization software that converts
descriptions of graphs in a simple text language either to diagrams in various formats
(e.g., images, SVG, PDF, Postscript) or for display in an interactive graph browser.
Using the SimpleXMLElement class (from the PHP library), the code loops through the
results of the SPARQL query <binding> elements, accessing the “name” attribute of
each one.
If the name attribute is ‘predicate’ (from the ?predicate variable in our SPARQL query),
the predicate is obtained from the <uri> subelement.
If the name attribute is ‘object’ (from the ?object variable in the query),
the value is obtained from the <uri> subelement or the <literal> subelement,
depending on which is present in the XML output of the query.
If the <uri> subelement is found, we use the uri target for display,
but use the complete uri for the link, for use in future queries that the user can
select.
The Graphviz text input file is constructed incrementally by adding Graphviz
format input statements for each <binding> element. Since Graphviz defines
links and colors in a separate section of the file, the XML <binding> elements
are searched a second time to create the Graphviz [label] entries. (To
improve performance, we may do this in a single loop populating two sections of the
Graphviz input at once.) A static set of predefined colors is defined in the module.
As new predicates (i.e., artist, associatedMusicalArtist, associatedBand,
musicComposer, producer, writer, etc.) are encountered, new colors are assigned by
taking the next one from the list and associating it with the new predicate.
Use of an associative array allows predicates of the same type to be assigned the
same color.
The Graphviz input string is completed when all the <binding> elements have
been processed and a header and footer are added.
The SVG data is created by sending the Graphviz input string to the
graphviz_filter_process method available by means of a Graphviz filter module
installed on the Drupal site. The SVG description is serialized by the filter
module. At this point the return information is structured using the Drupal function
drupal_json to create JSON formatted data. The SVG data and the XML
data are each added as named elements of the JSON data and returned to the caller.
To use this module, an HTML page with a short embedded JQuery script invokes the
module and displays the results. The HTML defines an input text field, a search
button and two <div> elements which initially contain no content. They are
given the IDs outputGraph and outputTable so content can be associated later using
JavaScript (JQuery).The user interface is a simple text field into which the user
types the name of the musician subject and then presses a search button which runs
a
JQuery function that posts the input field to the module previously described using
AJAX. The AJAX method waits for the data to be returned from the module in JSON
format. When the JSON data arrives, the returned SVG data and XML data are assigned
to the <div> elements with the correct IDs.
The figure below shows one of the Graphviz digraphs.
The SVG graphic produced by this Graphviz DOT file was too large to present here.
See Vocabutek Web Site.
Interface Screenshots
A sampling of our interface on Vocabutek Web Site is presented in the following figures.
As of this writing, the Graphviz visualization is expected to undergo changes.
Figure 12 illustrates how we entered queries in the Drupal interface to test.
Figure 12: Query Entry
Figure 13 displays the result of executing the query shown in Figure 12.
Figure 13: Query Result
We used the Drupal adminstration content management to organize queries into cascading
menus, as shown
in Figure 14.
Figure 14: Query Administration - Menu Organization
The result of the query organzation is shown in Figure 15
Figure 15: Cascading Menus
Figure 16 illustrates the interface for running a query to send to Graphviz for
rendering in SVG. Pre-tested queries can be copy/pasted into a single string (topmost
text field) or new
queries can be entered. The subject defaults to Eric Clapton but any musician can
be entered.
Figure 16: Visualize a Query - Copy/Paste or Enter
Finally, Figure 17 depicts part of the SVG resulting from a query, as
rendered by Graphviz. (This is one aspect we plan to improve over time.)
Figure 17: Graphviz AssociatedBand SVG
Challenges and Results
Our results were hampered by various anomalies or complications discovered in the
data sources
especially in terms of semantics.
In this section, we present some of the problems we encountered and either how they
were solved or how they might be approached
in the future.
Properties of a Musical Artist
We next consider a different view of Clapton predicates.
The query below returns properties with an object value that contains the string "Eric
Clapton".
Figure 18: Find Predicates for Which Clapton is (all or part of) the Object
SELECT DISTINCT ?predicate WHERE {
?s ?predicate ?o.
FILTER regex(?o, "Eric Clapton").
} ORDER BY ?predicate
When we subsequently connected the DBpedia and MusicBrainz data sources as discussed
in
Bridging Data Sources below, we obtained a third view of Clapton's properties,
shown in figure Clapton's Properties Combined.
Along Comes the Association
The paramount requirement in determining an Eric Clapton Number is the ability to
unambiguously identify those musicians with whom he directly worked. Originally we
thought this would be rather straightforward. Properties that seemed relevant to
making that determination included:
The musicalguests property was eliminated because it relates only to variety shows such as
"The Late Show". The starring property also proved to be unreliable since multiple
musicians might have all been considered stars of a performance but might not have
played together.
One case in point is the The Concert for Bangladesh;
a query returned Klaus Voorman, Billy Preston,
Bob Dylan, Eric Clapton, George Harrison, Leon Russell, Ravi Shankar, and Ringo Starr.
While
most of these performers did share the stage at one time during the concert, Ravi
Shankar did not
perform with the others. Furthermore, members of the band Badfinger did play with
most of the
others, but they were not returned by the query although we note that Badfinger is
cited in the abstract of
the Wikipedia article in the list of stars of the supergroup. On the other hand,
if we look at the infobox
on the concert's Wikipedia page,
Badfinger is not listed as one of the stars.
We considered the associatedActs property since that is the term used in the infobox.
As seen in Table 1, the infobox on Clapton's Wikipedia page
lists these associated acts:
"The Yardbirds, John Mayall & the Bluesbreakers, Powerhouse, Cream, Free Creek, The
Dirty Mac,
Blind Faith, J.J. Cale, The Plastic Ono Band, Delaney, Bonnie & Friends, Derek and
the Dominos".
Our initial attempt to retrieve these artists was a little surprising. Our query
was:
SELECT DISTINCT ?who
{ <http://dbpedia.org/resource/Eric_Clapton> <http://dbpedia.org/property/associatedActs> ?who.
} ORDER BY ?who
The result, shown below, was one long string
[6]
the contents of which represented a superset of the associated acts on the
Wikipedia page. The additional associated acts are underlined below.
"The Yardbirds, John Mayall & the Bluesbreakers, Powerhouse, Cream, Free Creek, Dire Straits,
George Harrison, The Dirty Mac, Blind Faith, Freddie King, Phil Collins,
J.J. Cale, The Plastic Ono Band, Delaney, Bonnie & Friends, Derek and the Dominos,
T.D.F., Jeff Beck, Paul McCartney, Steve Winwood, B.B. King, The Beatles, The Band"@en
It is unclear how there could be this disparity since the DBpedia data is derived
from Wikipedia. Although several of the underlined performers played with Clapton,
they were not in bands together (as far as we could determine). Furthermore, several
of the bands listed on the original Wikipedia page are arguably not exactly bands
in
the general sense since they existed only briefly and for a specific purpose. The
Dirty
Mac were a supergroup consisting of Eric Clapton, John Lennon, Keith Richards and
Mitch Mitchell that came together for The Rolling Stones' TV special. Free Creek was
a band composed of many musical artists, including Eric Clapton, Jeff Beck, Keith
Emerson, for one super-session album. Powerhouse only recorded a few songs, only three
of
which were released on a compilation album. These "bands" are certainly not
on a par with the others in the infobox since they were never intended to exist beyond
their stated purpose.
Perhaps our interpretation of the associatedActs property was incorrect? Further
examination of the DBpedia and Wikipedia documentation pointed us to the
template for the infobox of musicial artists,
shown on the right side of Table 1 in the earlier Wikipedia section.
The template description of associated_acts
from Wikipedia follows. (Numbers have been added for ease of reference and
the text has been reformatted.)
This field is for professional relationships with other musicians or bands
that are significant and notable to this artist's career.
This field can include, for example, any of the following:
a1) For individuals: groups of which he or she has been a member
a2) Other acts with which this act has collaborated on multiple occasions,
or on an album, or toured with as a single collaboration act playing together
a3) Groups which have spun off from this group
a4) A group from which this group has spun off
Separate multiple entries with commas.
The following uses of this field should be avoided:
b1) Association of groups with members' solo careers
b2) Groups with only one member in common
b3) Association of producers, managers, etc. (who are themselves acts)
with other acts (unless the act essentially belongs to the producer,
as in the case of a studio orchestra formed by and working exclusively
with a producer)
b4) One-time collaboration for a single, or on a single song
b5) Groups that are merely similar
Based on (a1), (a2) and (b4), it would seem that Powerhouse and arguably The Dirty
Mac and Free Creek
should be eliminated from the list of groups of which Clapton was a member.
[7]
Further investigation brought us to the numerous
DBpedia Infobox Mappings
and specifically to the ontology mapping for the Infobox for Musical Artists.
The (infobox) template property Associated_acts maps to two ontology properties,
associatedBand and associatedMusicalArtist, meaning that the infobox property
can refer to either a group (case (a3) and (a4) above) or to an individual (cases
(a1) and (a2)).
Therefore we turned our attention to queries involving the ontology properties
associatedBand and associatedMusicalArtist. The following query
asks for the bands in which Clapton was a member.
SELECT DISTINCT ?who WHERE
{<http://dbpedia.org/resource/Eric_Clapton> <http://dbpedia.org/ontology/associatedBand> ?who.
} ORDER BY ?who
The results are shown below. Note that the results are nearly identical to the string
superset
shown earlier. (The only exception is X-sample and a few band name variations.)
The same results are obtained if the predicate is replaced by http://dbpedia.org/ontology/associatedMusicalArtist.
The same results are obtained if the predicate is replaced by http://dbpedia.org/ontology/associatedMusicalArtist.
How are associatedBand, associatedMusicalArtist, and associatedActs related?
We believe the relationship varies across performers depending upon how individual
collaborators interpreted the terms.
Consider the following similar queries and their very different results, including
many performers who are
unfamiliar to all of the present authors.
Figure 20: Comparsion of associatedBand, associatedActs and associatedMusicalArtist
In order to address some of the more difficult questions or compare results between
data sources,
we needed a way to combine the RDF datasets, or at least to connect musical artist
identifiers
across DBpedia and MusicBrainz.
We employed the OpenLink Virtuoso Universal Server Virtuoso Universal Server as
an RDF store. While there are other very good RDF stores such as 4store, the Virtuoso Universal Server [8] is compelling because it supports reasoning. Backward-chaining
reasoning is the ability to derive new information based on existing
information at run time. This contrasts with forward chaining reasoning in which derived
information
is expressed explicitly. An example of forward chaining reasoning would be creating
and adding triples that represented new information that was inherent in information
the
datastore already contained.
[9]
There are two major costs to leveraging backward chaining reasoning. The
first and foremost cost is that derived information is created each time it is needed;
it is therefore quite likely
that the same information is derived repeatedly at run time. The advantage, however,
is that the derived information does not take up any space in the database. The
second requirement for leveraging backward chaining reasoning in Virtuoso is the use
of specific predicates, none of which were available in the native datasets with
which we worked.
Given that the IRI for Eric Clapton was different in DBPedia and MusicBrainz, we used
forward
chaining to express the relatedness between these two IRI’s and thereby facilitating
backward chaining reasoning in queries. We used the SPARQL
INSERT directive to create new statements which made
assertions explicit as a forward-chaining reasoning strategy. Specifically, the
assertions we added leveraged the owl:sameAs[10] predicates
Virtuoso defines to empower
reasoning. The
following query was used to create an assertion stating that the DBpedia Subject IRI
of type
http://dbpedia.org/ontology/MusicalArtist corresponds to the MusicBrainz Subject IRIs of type
http://musicbrainz.org/mm/mm-2.1#Artist whenever the DBpedia
rdf-schema#label matches the MusicBrainz dc:title
exactly.
This query resulted in the addition of 16,029 owl:sameAs assertions to our datastore which took
109,692,665 msec (30.47 hours) to complete.
Since neither these assertions nor the DBpedia and MusicBrainz assertions were in
the same NAMED GRAPH, it was necessary to use another unique
capability of the Virtuoso Universal Server called GRAPH GROUPs.
This enabled us to refer to multiple NAMED GRAPHs as if they were one. First we
created a GRAPH GROUP with the following command:
Then we placed the previously created owl:sameAs assertions, as well as
the NAMED GRAPHs containing DBpedia assertions and the MusicBrainz assertions
into the GRAPH GROUP:
To demonstrate that the owl:sameAs assertions we are indeed
intersecting the two datasets, consider the following query that determines the properties
connecting Eric Clapton to resources. Compare this to the figure
Clapton's DBpedia Properties - Version 1.
DEFINE input:same-as "yes"
SELECT DISTINCT ?predicate
FROM <http://group.dbpedia.inference> WHERE
{
?s ?predicate <http://dbpedia.org/resource/Eric_Clapton>.
} ORDER BY ?predicate
Any queries that benefit from using a single IRI to refer to the same artist in both
datasets
becomes available by adding DEFINE input:same-as "yes" before our result clause.
The same 30 results can be obtained with the UNION query below. However, this query
requires knowledge of the GUID-based MusicBrainz IRI for each musician of interest,
whereas the above query takes advantage of the previously established
correspondence between the methods of artist identification in DBpedia and MusicBrainz.
Therefore, the owl:sameAs approach is clearly the better solution.
SELECT DISTINCT ?predicate WHERE {
{
?s ?predicate <http://dbpedia.org/resource/Eric_Clapton>.
}
UNION
{
?s ?predicate <http://musicbrainz.org/mm-2.1/artist/618b6900-0618-4f1e-b835-bccb17f84294>.
}
} ORDER BY ?predicate
You Say It's Your Birthday?
Given the ability to bridge the two datasets, we can now issue queries that
compare data values across the sets. The query below obtains the birthdates for each
musical artist in both sources and returns the discrepancies.
A sampling of the nearly one thousand mismatched
birthdates follows. Typically the differences are one day, one month, or one year,
but note several major differences. We have yet to independently verify all of the
discrepancies, but we note that the birthday of George Harrison is incorrect in
MusicBrainz. In fact, MusicBrainz often uses "00-00" dates indicating only the year
is known. It is probably safe to conjecture that Courtney Love is not 21 and Five
For Fighting (singer-songwriter John Ondrasik) is not 14, so these birthdates are
more likely when their musical careers started. Wikipedia confirms that 1997 is indeed
the first "year active" for Five For Fighting, but Love's initial "year
active" is 1982. Clearly the interpretation of MusicBrainz's beginDate property
varies across artists.
Our tentative conclusion is that DBpedia is more accurate than
MusicBrainz with respect to birthdates.
musician DBPdate MBdate
50 Cent 1975-07-06 1976-07-06
Andy Partridge 1953-11-11 1953-12-11
Astrud Gilberto 1940-03-29 1940-03-30
Blind Willie Johnson 1897-01-22 1902-00-00 [5 year difference]
Blind Willie McTell 1898-05-05 1901-05-05 [3 year difference]
Carl Perkins 1932-04-09 1928-08-16 [4 year difference]
Courtney Love 1964-07-09 1990-00-00 [extremely different!]
David Lee Roth 1954-10-10 1955-10-10
Eddie Van Halen 1955-01-26 1956-01-26
Edgar Winter 1947-12-28 1946-12-28
Five for Fighting 1965-01-07 1997-00-00 [extremely different!]
Frankie Valli 1934-05-03 1937-05-03 [3 year difference]
George Harrison 1943-02-25 1943-02-24
Jennifer Lopez 1969-07-24 1970-07-24
Top Record Labels
One of our original questions was "Is there a predominant record label in the music
world?"
The following query answers that question.
SELECT ?label (count(?label) as ?count)
{
?artist <http://dbpedia.org/ontology/recordLabel> ?label
}
ORDER BY DESC (?count) LIMIT 20
In order to determine which artist is most influential in rock music, we needed to
be able to
reliably specify the genre of interest. However, formulating queries involving genres
is more difficult than it would seem. DBpedia (and presumably Wikipedia)
defines 2,887 musical genres of which 330 contain "rock" in their label. Rock-related
genres literally run the gamut
from A to Z -- Aboriginal_rock to Zulu_rock (really!). To our surprise, the single
concept of "rock and roll music"
is represented by 18 distinctly different IRIs, as determined by the following query.
[11]
For any given album, more than one genre may apply. For example, the 1975 album
"There's One in Every Crowd" is classified as both reggae and blues-rock.
For those wondering which Clapton album could possibly be considered jazz or orchestral,
that distinction
belongs to the first "Lethal Weapon" soundtrack, which is also designated as blues.
However, the genre query above does not capture all genres associated with Clapton.
For example,
the genre for "Lethal Weapon 3" soundtrack is simply "soundtrack".
If we wish to display all of Clapton's blues and rock albums, we could use a query
such as:
SELECT DISTINCT (?genre) ?album WHERE
{
?album <http://dbpedia.org/ontology/artist> <http://dbpedia.org/resource/Eric_Clapton>.
?album <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/ontology/Album>.
?album <http://dbpedia.org/ontology/genre> ?genre .
FILTER ( regex (?genre, "[Rr]ock") || regex (?genre, "[Bb]lues"))
}
ORDER BY ?album
To display all of Clapton's albums and their associated genre(s), we used the following
query:
SELECT ?album ?genre WHERE
{
?album <http://dbpedia.org/ontology/artist> <http://dbpedia.org/resource/Eric_Clapton>.
?album <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/ontology/Album>.
?album <http://dbpedia.org/ontology/genre> ?genre .
}
ORDER BY ?album
The result including albums with several genres listed follows. (The common portion
of the IRI,
http://dbpedia.org/resource/ has been removed from each resource to make the results
more readable.)
However, if we omit a reference to genre and simply ask for Clapton's albums
using the query:
SELECT ?album WHERE
{
?album <http://dbpedia.org/ontology/artist> <http://dbpedia.org/resource/Eric_Clapton>.
?album <http://www.w3.org/1999/02/22-rdf-syntax-ns#type> <http://dbpedia.org/ontology/Album>.
}
ORDER BY ?album
then the results contain 7 additional albums (below). We have not been able to determine
the
reason for this disparity since there is genre information for each of them when we
dereference
their IRIs on the DBpedia site.
Clapton_%281973_album%29 dbpprop:genre Rock
[compilation from 1973]
Complete_Clapton dbpprop:genre Blues
[compilation from 2007]
Edge_of_Darkness_%28soundtrack%29 dbpprop:genre Soundtracks
[18 minute 1985 soundtrack for British TV series]
Eric_Clapton_at_His_Best dbpprop:genre Rock
[compilation from 1972]
Lethal_Weapon_3_%28soundtrack%29 dbpprop:genre Orchestral, Jazz, and Blues
[Wikipedia soundtrack entry shares page with the movie; 2 infoboxes]
Live_from_Madison_Square_Garden_%28Eric_Clapton_and_Steve_Winwood_album%29 dbpprop:genre Blues/Rock
[note the slash]
Rush_%28soundtrack%29 dbpedia-owl:type dbpedia:Soundtrack
[Wikipedia soundtrack entry shares page with the movie; but no album infobox]
A better solution for managing the permutations between the
representations for the genre concept "Rock and Roll" as well as its narrowings would
be to
again use reasoning. The semantically equivalent variant genres could be represented
as
rdfs:subProperties of one
another, thereby enabling a single genre representation to refer to many. [12] This would
require analogous steps to where we created and loaded assertions, defined a graph,
define a graph group, and in this case define a rule_set in Virtuoso. Then, using the
proper syntax we could query using that rule_set to include exactly the genres we
intend to use in our queries. A less elegant approach would be to enumerate them one
by one and construct a
UNION of results.
Limitations and Further Efforts
We recognize several limitations in our work to date:
data inconsistencies
need for further RDF visualization work
need to address more of the original problems
constructing queries across datasets
Each is discussed in the following subsections.
Data Inconsistencies
While we would wish the source RDF data could be regarded as ground truth, we realize
there are
several problems with our sources.
Errors of omission: Since the original data in Wikipedia is community-entered,
it is predictable that certain facts will be missing from the data but are in fact
true. Some
of thes facts may be obscure while others may be more obvious to a subject matter
expert
for the given topic. For example, what if Eric Clapton were to enter his own data
about himself?
Poor data curation: In some cases, Wikipedia data may
have been present and complete, but there might have been problems in the extraction
process from Wikipedia
to DBpedia.
Erroneous data: Such problems are due to simple data entry errors,
unintentional errors in stating what the data entry person considers facts,
or possibly intentional falsehoods or unsubstantiated facts.
Unclear semantics: The various properties which were of importance to us were
generally not defined in the ontology, as discussed in Working Definitions. We struggled to interpret
these terms in a consistent manner. It is possible that individuals who care less
about the precision of
information may have entered relationships not necessarily using the correct semantics.
We believe that the ambiguity of the musician-related properties was the most significant
problem with the reliability of the data and therefore the biggest challenge in testing
our
hypothesis.
Further RDF Visualization Work
Since AJAX facilitates updating a web page without reloading the entire page,
we plan to insert additional links within the SVG graphic to add more
interactivity, invoking additional queries. Links could also be added during the
module processing. CSS and XSLT could also be used to enhance the XML
presentation.
The current visualization graph is quite wide and long, making it difficult to
view in a web browser without additional panning and zooming capabilities. Useful
visualization of the result set is difficult but we intend to improve the
visualization to facilitate traversing the dataset using hyperlinks. The
hyperlinks are there now, but there is currently a problem with the JavaScript
POST method.
Further Attempts to Address the Original Problems
We initially formulated nearly two dozen questions [see Problems]
that we believed we could use the DBpedia and MusicBrainz
RDF datasets to answer. As of this July writing, few of these questions have been
answered definitively.
If properly bounded, the following are the questions we should be able to address
in the future.
Which recording artist has directly played with the most musicians?
Which recording artist has the most connections within six degrees?
Which musician has been a session man for the most number of artists?
Which recording artist was most active during a particular decade?
Among all artists of a particular genre, who has played with the most other musicians?
Which rock artist's extended graph has the most other artists in 2 degrees? 3 degrees?
4 degrees?
Who has appeared on the most albums?
Which musician-related properties are reversible (inverse makes sense)?
Who created the most songs?
Which song has been recorded the most times by any artists?
("Yesterday" and "White Christmas" are typically cited.)
What is the average age of a musician when he/she first joined a band?
Other questions we may potentially be able to address include:
If we weight results by the length of time a band stays together, how does that impact
other queries?
Does total number of songs or album released correlate with other measures of success?
Which solo artist has had the longest career?
Which band has been together (in some form) the longest time?
For bands with changing membership, can we conclude which configuration lasted the
longest?
What is the "Eric Clapton number" (a la Kevin Bacon number) for various musicians?
Queries Across Datasets
Since we have established a mechanism to
refer to a given artist in both datasets using a single IRI,
we are prepared to ask queries that span the datasets, including
queries that would not be possible without both sources. A few such queries are as
follows:
Which albums of a given artist occur in one dataset but not the other?
Since we are using Virtuoso which does not support the SPARQL 1.1 MINUS construct,
we would need to use OpenLink's proprietary approach. [13]
Which albums that occur in both datasets have the same title and record producer but
different release dates?
This might be difficult to determine because of differences in representation of the
record producer name.
For a given artist, return their spouse(s) and children even if they themselves were
not
musical artists. MusicBrainz only has names of children if they themselves are artists
(e.g., has
John Lennon's sons but not Paul McCartney's children).
Put the LOD Right on Me
Regrettably, we focused most of our attention on DBpedia, MusicBrainz and Freebase
web sites, rather than
casting a broader net to the Linking Open Data
community effort initiated by the W3C's (now closed) Semantic Web Education and Outreach
(SWEO) Interest Group.
As we were finalizing this paper, we discovered several key
resources that could prove extremely useful in answering some of our original
questions; we hope to explore resources such as the following over the next year:
Next we offer our subjective opinion about the relative advantages and disadvantages
of our approach
as compared to simply using the DBpedia SPARQL endpoint
(which also runs on a Virtuoso server).
Advantages of Our Approach
Our Drupal frontend to a dedicated Virtuoso server under our fine-grained control
afforded us several advantages
[14], namely:
No execution timeout - Time-intensive queries were allowed to run
to completion rather than being subject to an externally specified time limit on a
shared server.
Higher limits - SPARQL endpoints shared by multiple users are
often constrained to a limited number of results (i.e., 2000 for DBpedia endpoint).
We did not
need to limit the number of results returned since our number of "users" never exceeded
three.
Direct access to raw RDF at command line level - When results
were not as expected or did not seem to make sense, the ability to examine the actual
RDF files
(which had been ingested into our datastore) in our Linux environment was invaluable.
Collaborative SPARQL development for saving queries in Drupal -
As the three authors worked remotely, it was quite helpful to enter our SPARQL queries
into
a Drupal interface, run the queries, save the queries, and optionally enter them into
a cascading
menu system for each other to try so we could discuss results.
More precise SPARQL timing metrics - Again, since we could
control which other processes were running and other variables that would otherwise
impact
timing metrics, we could more accurately time our queries. For example, we found it
took
on average 3477.6 msec for a query with one FILTER statement comprised of two
regex expressions joined by "&&" compared to only
731 msec for a similar query in which each regex expression was the parameter of
a separate FILTER statement. See the first example in the Musical Genres section.
Test harness for repeatedly running the same query - We employed
a perl test harness for iterative execution and averaging of results. For example,
the previously
mentioned FILTER and regex tests were run 1,000 times.
Examination of error messages - We had access to server logging
of error messages.
Disadvantages of Our Approach
Static datastore - After we ingested the RDF files, our datastore
did not change so it could not take advantage of possible improvements in the DBpedia
and Musicbrainz
data throughout the first half of 2011. On the other hand, this was also an advatange
since it meant
we were not subject to subtle changes that might have impacted previous queries.
Limited knowledge of inner workings - Since we were not members
of either the DBpedia or Musicbrainz developer communities, we were not privy to any
internal documentation
that might have answered some of the questions we raised and helped us to address
some of the problems we
encountered. In a volunteer developer collaborative environment, it is likely that
user-facing
documentation may lag behind actual implementation changes; relevant examples may
also be lacking.
Recommendations
The problems we encountered suggest several recommendations we would like to share
with the DBPedia and MusicBrainz communities:
Publish example SPARQL queries: Although sample data is readily available
on the music sites, we did not find SPARQL queries, just endpoints to explore.
Provide detailed explanation of IRI conventions and property semantics: We
found that it was not always obvious what the significance was for IRI differences,
nor what non-alphabetic
characters should be used in multi-word IRIs. For example, what is the difference
between
http://dbpedia.org/ontology and http://dbpedia.org/property IRIs?
Include examples of intersecting graphs from other music-oriented sites:
In the spirit of Linking Open Data,
it would be helpful if both DBpedia and MusicBrainz created and posted
demos (or links to demos) on their sites to illustrate linking of their data sources
to other popular music data stores.
Conclusions
Although at this time we have fallen far short of our lofty goal of answering two
dozen complex questions
by means of RDF and SPARQL, we believe we have uncovered several significant challenges
regarding the consistency of the data sources and the interpretation of the semantics
underlying various key properties of musical artists. We believe similar problems
would be
encountered with using other large community-entered datasets of relatively low quality.
While Wikipedia editors often focus on crafting the most accessible presentation of
encyclopedic
information for human consumption, there are other factors to consider. Individuals
working with data at the
semantic level could improve the semantic representation by isolating outliers, as
in
the case of multiple representations of the "Rock and Roll music" genre concept.
Data quality and therefore utility would be greatly enhanced
by making the necessary modifications to "normalize" or decrease
spurious diversity that is not based on any actual or intended semantic distinction.
[15]
In the end, any accurate measure of impact a given musician has on a genre is likely
to be the product of a
weighted measurement of several variables and is therefore largely subjective. The
number
of bands someone played with, the number of songs they wrote, the number of albums
they
produced or the length of time they were active are only part of the equation. An
artist
who wrote a single song that has enjoyed frequent airplay for decades and/or was covered
by artists in many
different genres clearly demonstrates a substantial impact. Other measurements not
available in
the RDF datasets are likely to be even more revealing of an artist's true impact,
such as the
number of concerts they played, the number of people who recognize their name, and
their
success on music charts (i.e., Billboard).
By the same token, a strong argument could be made that an individual musician whose
career lasted only a short time with a limited repertoire but perhaps created or
influenced a new genre could be regard to have a major impact. Therefore, one
would have to somehow compare Eric Clapton to early genre pioneers. For example, Robert
Johnson, a major influence on Clapton, recorded only 29 songs and basically enjoyed
only
a 2-year (1936-37) recording career. Yet Clapton himself has called Johnson "the
most important blues singer that ever lived".[16]
Johnson was ranked fifth in Rolling Stone's list of 100 Greatest Guitarists of All Time.
Given our knowledge and supporting resources surrounding Eric Clapton's career, we
believe
that some of what we perceive as errors are indeed just that -- mistakes. Whether
those
mistakes are based on individuals simply being incorrect, or not understanding the
semantics really does not matter in the final analysis. What this highlights, for
us,
is a limitation induced
by the Open World Assumption.
When an assertion is not made, we cannot also assume its
negation. Having assertions stating that Clapton created 30 albums does not mean he
did
not create 45. Likewise, because we know that Clapton did participate in collaboration
efforts like Powerhouse, The Dirty Mac and Free Creek, we cannot state unequivocally
that
it happened only a
single time and therefore violates the semantics of the associatedActs predicate.
[Clapton FAQ]
Eric Clapton Frequently Asked Questions,
part of The Eric Clapton Fan Club Magazine. [online]. [cited 08 Apr 2011].
http://www.whereseric.com/eric-clapton-faq
[Feigenbaum and Prud'hommeaux 2008]
SPARQL By Example - A Tutorial. [online] [cited 08 Apr 2011].
Feigenbaum, Lee and Prud'hommeaux, Eric. Cambridge Semantics, 2008. Updated 2011-01-25.
http://www.cambridgesemantics.com/2008/09/sparql-by-example/
[1] He believes it was circa 1972 in Buffalo, New York. But his recollection
of the Seventies is much like his recollection of the Sixties.
[2]
According to Whitburn 2006,
Clapton released 41 albums (including reissues) from 1970 to 2005. (He has released
three more as of July 2011.)
This includes eleven Top 10 albums
and three number 1 albums. Whitburn devised a formula for ranking artists by their
chart success,
with rankings per decade and overall. According to Whitburn's calculations as of 2006,
Clapton ranked #10 in the 1990's and ranked #12 all-time; he ranks #21 in most charted
albums.
[3] IRI is Internationalized Resource Identifier, a generalization of the Uniform Resource
Identifier (IRI)
enabling the use of Unicode. In this paper, we refer to IRI rather than URI since
SPARQL
technically permits IRIs.
[4] If we add one more clause to the query asking also for
http://dbpedia.org/resource/Category:English_rock_guitarists,
then Eric Clapton is the only result.
[5] The full Freebase data set in TSV format is 1.3 GB compressed. The entire
Freebase data dump is also available as quadruples in a 3.36 GB download.
When uncompressed, the single file freebase-datadump-quadruples.tsv is 27.3
GB.
[6] Technically this is an object value comprised of a number of strings concatenated
into a single field.
[8] The OpenLink Virtuoso Universal Server supports relational data
management, XML and RDF data management, free text content management with
full text indexing and web services, and functions as a document web server,
linked data server and web application server.
[9] "Inference is moreso the means or mechanisms by which reasoning
occurs. Reasoning is the 'goal' whereas inferencing is the
'implementation'". -- From answers.semanticweb.com.
[10]owl:sameAs is often used in establishing mappings between two or more ontologies.
[11] Actually, there is a 19th IRI if the query is not constrained to require the string
"roll":
http://dbpedia.org/resource/Rock_music. This link redirects to a page which among other things lists
many subgenres of rock.
[12] However, we chose to use rdfs:subClassOf because it was the most appropriate RDFS property that the
Virtuoso database reasoner supported.
[13] It is possible that NOT EXISTS in SPARQL 1.0 would be sufficient.
