=Paper=
{{Paper
|id=None
|storemode=property
|title=A Method for Obtaining Semantic Facets of Music Tags
|pdfUrl=https://ceur-ws.org/Vol-633/wom2010_paper4.pdf
|volume=Vol-633
}}
==A Method for Obtaining Semantic Facets of Music Tags==
https://ceur-ws.org/Vol-633/wom2010_paper4.pdf
A Method for Obtaining Semantic Facets of Music Tags
Mohamed Sordo Fabien Gouyon Luís Sarmento
Universitat Pompeu Fabra INESC Porto LIACC/FEUP, Univ. do Porto
Barcelona, Spain Porto, Portugal Porto, Portugal
mohamed.sordo@upf.edu fgouyon@inescporto.pt las@fe.up.pt
ABSTRACT However, there exists no consensual taxonomy for music.
Music folksonomies have an inherent loose and open seman- Previous research showed the music industry uses inconsis-
tics, which hampers their use in structured browsing and tent taxonomies [6], even when restricting to a single and
recommendation. In this paper, we present a method for widespread facet such as the music genre. Also, expert-
automatically obtaining a set of semantic facets underly- defined taxonomies (music-related or not) have two funda-
ing a folksonomy of music tags. The semantic facets are mental problems. First, they are very likely to be incom-
anchored upon the structure of the dynamic repository of plete, since it is impossible for a small group of experts to
universal knowledge Wikipedia. We illustrate the relevance incorporate in a single structure all the knowledge that is
of the obtained facets for the description of tags. relevant to a specific domain. Second, since domains are
constantly evolving taxonomies tend to become quickly out-
dated –in music, new genres and techniques are constantly
Categories and Subject Descriptors emerging.
H.3.1 [Information Storage and Retrieval]: Content An alternative strategy for describing music consists in
Analysis and Indexing—Dictionaries, Linguistic processing; relying on the broadness of the web and making use of
H.5.5 [Information Storage and Retrieval]: Sound and the “wisdom of the crowds”. Many music websites allow
Music Computing users themselves to assign their own descriptive tags to mu-
sic items (artists, albums, songs, playlists, etc.). For in-
General Terms stance, users of the website Last.fm tagged the band Radio-
head as “90s”, “00s”, “alternative”, “post-punk”, “britpop”,
Algorithms, Experimentation, Languages “best band ever”, among other things. The combination of
annotations provided by thousands of music users leads to
Keywords the emergence of a large body of domain-specific knowledge,
Music tagging, Last.fm, Wikipedia, Social music usually called folksonomy. Due to its informal syntax (i.e.
direct assignment of tags), the tagging process allows the
collective creation of very rich tag descriptions of individual
1. INTRODUCTION music items.
Music is a complex phenomenon that can be described When compared to taxonomies defined by experts, music
according to multiple facets. Descriptive facets of music are folksonomies have several advantages. First, completeness,
commonly defined by experts (e.g. stakeholders in the music they ideally encompass all possible “ways to talk about mu-
industry) in professional taxonomies. Multifaceted descrip- sic”, including both lay and expert points of view. Second,
tions are especially useful for music browsing and recom- due to the continuous nature of the tagging process, folk-
mendation. For instance, recommendations of the Pandora sonomies tend to be well updated. Third, they usually in-
Internet radio use around 400 music attributes grouped in corporate both commonly accepted and generic concepts, as
20 facets,1 as for instance Roots (e.g. “Afro-Latin Roots”), well as very specific and local ones.
Instrumentation (e.g. “Mixed Acoustic and Electric Instru- It seems reasonable to assume that folksonomies tend to
mentation”), Recording techniques (e.g. “Vinyl Ambience”), encompass various groups of tags that should reflect the un-
or Influences (e.g. “Brazilian Influences”). derlying semantic facets of the domain including not only
1
http://en.wikipedia.org/wiki/List_of_Music_ traditional dimensions (e.g. instrumentation), but also more
Genome_Project_attributes subjective ones (e.g. mood). However, the simplicity and
user-friendliness of community-based tagging imposes a toll:
there is usually no way to explicitly relate tags with the cor-
responding music facets. For instance, a user may assign a
number of tags related with music genre without ever actu-
ally explicitly specifying that they are about “music genre”.
WOMRAD 2010 Workshop on Music Recommendation and Discovery, For providing a flexible browsing experience, this is a sig-
colocated with ACM RecSys 2010 (Barcelona, SPAIN) nificant disadvantage of folksonomy-based classification in
Copyright c . This is an open-access article distributed under the terms
of the Creative Commons Attribution License 3.0 Unported, which permits
relation to classification based on taxonomies, where the in-
unrestricted use, distribution, and reproduction in any medium, provided formation about which facets are being browsed can be made
the original author and source are credited.
explicitly available to the user. by maximizing the overlap between Wikipedia pages and a
In this paper, we approach an essential research question list of frequent tags from the folksonomy. As the resulting
that is relevant to bridging this gap: Is it possible to auto- network still represents a very large number of nodes, in a
matically infer the semantic facets inherent to a given music second step, we focus on the most relevant ones (node rele-
folksonomy? A related research question is whether it is vance being defined as an intrinsic property of the network).
then possible to classify elements of that music folksonomy This step also includes additional refinements.
with respect to the inferred semantic facets?
We propose an automatic method for (1) uncovering the 3.1 Obtaining a Music-Related Network
set of semantic facets implicit to the tags of a given mu- Wikipedia pages are usually interlinked, and we use the
sic folksonomy, and (2) classify tags with respect to these links between two particular types of pages (i.e. articles
facets. We anchor semantic facets on metadata of the semi- and categories) to construct a music-related network. Con-
structured repository of general knowledge Wikipedia. Our cretely, we use the DBpedia knowledge base (http://dbpedia.
rationale is that as it is dynamically maintained by a large org/) that provides structured, machine-readable descrip-
community, Wikipedia should contain grounded and updated tions of the links between Wikipedia pages (DBpedia uses
information about relevant facets of music, in practice. the SKOS vocabulary, in its 2005 version).2 In particular,
we make use of two properties that connect pages in DBpe-
dia: (1) the property subjectOf, that connect articles to cate-
2. RELATED WORK gories (e.g. the article “Samba” is a subjectOf of the category
Music tags have recently been the object of increasing “Dance music”, and (2), the property broaderOf, that con-
attention by the research community [3, 4]. A number of nect categories in a hierarchical manner (e.g. the category
approaches have been proposed to associate tags to music “Dance” is a broaderOf of the category “Dance music”, which
items (e.g. a particular artist, or a music piece) based on an is a broaderOf of the category “Ballroom dance music”).
analysis of audio data [1, 9], on the knowledge about tag co- We start from the seed category “Music” and explore its
occurence [5], or on the extraction of tag information from neighbourhood from the top down, checking whether con-
community-edited resources [8]. However, in most cases, nected categories can be considered relevant to the music
such approaches consider tags independently, i.e. not as el- domain. A category is considered relevant if it satisfies any
ements in structured hierarchies of different music facets. of the two following conditions:
When hierarchies of facets are considered, they are usu-
ally defined a priori, and greatly vary according to authors. • It is a tag from the folksonomy, such as for example
For example, [4] groups tags in the following facets: genre, “Rock and Roll”. (This condition will be referred to as
locale, mood, opinion, instrumentation, style, time period, isMusical );
recording label, organizational, and social signaling.
To our knowledge, however, few efforts have been dedi- • At least one of its “descendants” is a tag from the folk-
cated to the particular task of automatically identifying the sonomy and the substring “music” is included in the
relevant facets of music tags. In their work on inferring title or the abstract of the corresponding Wikipedia
models for genre and artist classification, Levy et al. apply article. (This condition is further referred to as is-
dimensionality reduction techniques to a data set of tagged TextMusical.)
music tracks in order to obtain their corresponding com- The “descendants” of a category are fetched from DBpe-
pact representations in a low-dimensional space [5]. They dia using the two connecting properties previously described.
base their approach on tag co-occurrence information. Some These descendants can be either “successors” (i.e. all direct
emerging dimensions can be associated to facets such as Era subjectOf and broaderOf of this category), or successors of
(e.g. the dimension [90s]). However, most of the dimen- successors, and so on. This iterative search is limited by a
sions thus inferred are, in fact, a combination of diverse mu- maximum depth, empirically fixed to a value of 4. Indeed,
sic facets, such as for example the dimension [guitar; rock], experiments with smaller values yielded a significant reduc-
which includes concepts of instrumentation and of genre. tion of the tag coverage, while experiments with greater val-
Cano et al. use the WordNet ontology to automatically ues did not increase significantly the coverage.
describe sound effects [2]. Albeit the very large amount of If any of the previous conditions is satisfied, the cate-
concepts in WordNet, they report that it accounts for rela- gory, its successors and their edges are added to the net-
tively few concepts related to sound and music, and propose work. Otherwise, the category and all incident edges are
an extension specific to the domain of sound effects. On removed. The algorithm proceeds iteratively (following a
the one hand, they illustrate that browsing can indeed be Breadth-First search approach) until no more categories can
greatly enhanced by providing multifaceted descriptions of be visited. A summarized version of the method for obtain-
items. On the other hand however, it is our belief that, ing a music-related network is described in algorithm 1.
because of their necessary stability, existing ontologies are
not the most adapted tool to describe domains of knowledge 3.2 Finding Relevant Facets
with inherent open and dynamic semantics, such as music. Once the network of music-related categories is built, the
next step is to find the nodes that are potentially more rel-
3. METHOD evant to the network than others.
Our method consists in using metadata from Wikipedia to We invert the direction of the edges of the network in
infer the semantic facets of a given music folksonomy. This order to point back in the direction of the most generic cat-
is performed in two steps. In the first step, we specialize the egory, i.e. “Music”, and we compute the PageRank of the
very large network of interlinked Wikipedia pages to the spe- 2
http://www.w3.org/TR/2005/
cific domain of the music folksonomy at hand. This is done WD-swbp-skos-core-spec-20051102/
Data: C = ∅, a list of categories (a queue, initially
empty); N = (V, E), a directed network with a
set of nodes V and a set of edges E (initially
empty);
Result: N , network with music nodes;
C ← C ∪ “M usic”;
while C 6= ∅ do
c ← f irst element of C;
C ← C − c;
if (c isMusical) ∨ ((at least one descendant of c
( ∧ (c isTextMusical)) then
isMusical)
V ← V ∪ c ∪ successors(c)
N
E ← E ∪ edges between c and successors(c)
C ← C ∪ successors(c)
else ( Figure 1: Example of subnetwork in our data. Dot-
V ←V −c ted lines correspond to Wikipedia categories that
N are also Last.fm tags. Dashed lines correspond to
E ← E − all edges incident in c
categories not kept. Plain lines correspond to facets
end
kept.
end
Algorithm 1: Pseudo-code for the creation of a network
of music-related categories from Wikipedia. 4. RESULTS
We experimented our method on a large dataset of artist
tags, gathered from Last.fm during April 2010. The dataset
resulting network. PageRank [7] is a link analysis algorithm
consists of around 600,000 artists and 416,159 distinct tags.
that measures the relative relevance of all nodes in a net-
This dataset was cleaned in order to remove noisy/irrelevant
work. In PageRank, each node is able to issue a relevance
data: (1) tags were edited in order to remove special char-
vote on all nodes to which it points to (thus the need for re-
acters such as spaces, etc.; (2) tags were filtered by weight3 ,
orienting the edges). The weight of the vote depends on the
only tags with a weight ≥ 1 were kept; and (3) tags were fil-
relevance of the voting node (i.e. relevant nodes issue more
tered by popularity, keeping only tags with popularity ≥ 10,
authoritative votes). The process runs iteratively, and (un-
i.e. keeping only tags that were assigned to at least 10
der certain conditions) converges to a stable relative ranking,
artists. As a result, the final dataset consists of 582,502
where nodes to which more edges from other relevant nodes
artists, 39,953 distinct tags, and 9.03 tags per artist.
converge (directly or indirectly) are considered more rele-
After running both stages of our method, we obtained a
vant. For initializing the PageRank algorithm, we set the
list of 333 candidate facets. Table 1 contains the top-50
initial weight of each node to 0.
facets, ordered by pagerank (top to bottom, left to right).
In order to capture general yet complementary facets of
music, we aim at reducing semantic overlap as much as pos-
sible by applying the following filters: Table 1: Top-50 Wikipedia music facets
Music genres Aspects of music
Stub Filter: We remove all categories with substring “ by ” Music geography Hip hop genres
Musical groups Music of California
and “ from ”. We noticed that many categories in Music industry Music theory
Wikipedia are actually combinations of two more gen- Musicians
Musical culture
Rock and Roll Hall of Fame inductees
Musical subcultures
eral categories, as for instance “Musicians by genres”, Occupations in music Recorded music
Music people Musical quartets
which is halfways between “Musicians” and “Music genres” Record labels Music festivals
Music technology East Asian music
(see also figure 1). Further, we also remove categories Sociological genres of music Centuries in music
that include “ music(al) groups” (e.g. “Musical groups- Music publishing companies
Musical instruments
Musical composition
Musical quintets
from California” that has hundreds of connected cat- Anglophone music Southern European music
Music of United States subdivisions Music software
egories, hence a high PageRank). Most of these cate- Western European music Incidental music
American styles of music Years in music
gories are used as stubs, even sometimes explicitly so Radio formats Music websites
we also excluded categories with the word “stub”. Music publishing
Albums
Guitars
Music competitions
Musical techniques Musical eras
Over-Specialization Filter: We exclude all categories that Wiki music Music and video
Music history Musical terminology
include lexically a more relevant category. Many rel- Music performance Music halls of fame
Music publishers “people” Dates in music
evant categories are specializations of other more rel-
evant ones, this occurs mostly with concepts related
to anglophone music, which are described in great de-
tail in Wikipedia (e.g. “American Musicians” includes
4.1 Assigning facets to tags
“Musicians” that has a higher PageRank). In order to assign a set of facets to a given Last.fm tag, we
process the subnetwork of Wikipedia pages specialized to the
Tag Filter: We remove all categories that are tags. Our Last.fm folksonomy (obtained in section 3.1), as described
objective is to uncover music facets that are implicit in algorithm 2 (Note that this process is restricted to tags
to the tags that make up a folksonomy. In general, tags that can be matched to one of the nodes in the network).
are elements of such facets, not the facets themselves. 3
i.e. Last.fm “relevance weight”, which goes from 0 to 100
Table 2: Sample of the top tags for various music facets inferred
Music genres Occupations in music Musical instruments Aspects of music
Sufi music Troubadour Melodica Rhythm
Dance music Bandleaders Tambourine Melody
Indietronica Pianist Drums Harmony
Minimalism Singer-songwriter Synthesizers Percussion
Singer-songwriter Flautist Piano Chords
Music software Music websites Music competitions Musical eras
Nanoloop Mikseri.net Nashville Star Baroque music
Scorewriter PureVolume American Idol Ancient music
MIDI Allmusic Melodifestivalen Romantic music
DrumCore Jamendo Star Search Medieval music
Renoise Netlabels Eurovision Song Contest Renaissance music
Data: C = ∅, a list of categories (initially empty); F , a the underlying semantic facets of the Last.fm folksonomy,
list of top-N music facets; t, a Last.fm tag; using Wikipedia as backbone for semi-structured semantic
Result: T F , list of facets applied to tag t; categories.
iter ← 1; There are many avenues for future work. First and fore-
T F = ∅; most, we intend to evaluate the relevance of the obtained
while (F 6= ∅) ∨ (iter ≤ maxIter) do facets via systematic evaluations of tag classification. We
C ← C ∪ predecessors(t); will also study the distributions of music facets with respect
if (∃f ∈ (F ∩ C)) then to artist popularity. Further work should also relate to eval-
TF ← TF ∪ f uating the usefulness of the obtained facets in a number of
F ←F −f tasks, such as music recommendation, or tag expansion. We
end also intend to release the data (and code used to obtain it)
iter ← iter + 1 in order to stimulate its use by fellow researchers.
end
Algorithm 2: Pseudo-code for assigning Wikipedia facets
to Last.fm tags
6. ACKNOWLEDGMENTS
Thanks to Òscar Celma (BMAT), Eduarda Mendes Ro-
drigues (FEUP) and anonymous reviewers for useful com-
Given a Last.fm tag t, we look at its “predecessor” cate- ments. This work was partly supported by the Ministerio
gories c, or more formally: de Educación in Spain, and the Fundação para a Ciência e
a Tecnologia (FCT) and QREN-AdI grant for the project
predecessors(t) = {c|(t broaderOf (c)) ∨ (t subjectOf (c))}. Palco3.0/3121 in Portugal.
If any of these predecessors is a top-N facet, it is then as-
signed to t. The process continues iteratively until no more 7. REFERENCES
facets can be assigned to the tag, or a maximum number [1] T. Bertin-Mahieux, D. Eck, F. Maillet, and P. Lamere.
of iteration (maxIter) is exceeded. We empirically set this Autotagger: A model for predicting social tags from
value to 8. This condition can be interpreted as the maxi- acoustic features on large music databases. JNMR,
mum distance in the network between a tag and a facet. 37(2):115–135, 2008.
Table 2 presents a small subset of the obtained facets, [2] P. Cano, M. Koppenberger, P. Herrera, O. Celma, and
followed by a subset of their corresponding list of top tags. V. Tarasov. Sound effect taxonomy management in
Top tags are chosen based on the distance (in number of production environments. In AES, 2004.
successive edges in the music network) to the given facet. [3] O. Celma. Music Recommendation and Discovery - The
The relevance Rtf of a music facet f to a tag t is com- Long Tail, Long Fail, and Long Play in the Digital
puted as the normalized inverse distance dtf –in number of Music Space. Springer, 2010.
successive edges– between t and f : [4] P. Lamere. Social tagging and Music Information
1 Retrieval. JNMR, 37(2):101–114, 2008.
dtf
Rtf = P 1 [5] M. Levy and M. Sandler. Learning latent semantic
i dti models for music from social tags. JNMR,
For example, in figure 1, given the tag bulgarian hip-hop, 37(2):137–150, 2008.
our method starts navigating through the predecessors of [6] F. Pachet and D. Cazaly. A taxonomy of musical
this tag until finally reaching two music facets: Music genres genres. In RIAO, 2000.
and Music geography: [7] L. Page, S. Brin, R. Motwani, and T. Winograd. The
pagerank citation ranking: Bringing order to the web.
bulgarian hip-hop: {(Music_genres, 0.4), Technical report, Stanford InfoLab, 1999.
(Music_geography, 0.6)} [8] L. Sarmento, F. Gouyon, and E. Oliveira. Music artist
tag propagation with wikipedia abstracts. In
5. SUMMARY AND FUTURE WORK ECIR-WIRSN, 2009.
Although potentially more complete and up-to-date than [9] D. Turnbull, L. Barrington, D. Torres, and
taxonomies, music folksonomies lack structured categories, G. Lanckriet. Semantic annotation and retrieval of
a particularly relevant aspect to browsing and recommenda- music and sound effects. IEEE TASLP, 2(16):467–476,
tion. In this paper, we addressed the problem of uncovering 2008.