Information Technology Reference
In-Depth Information
compared with the documents in the collection
using approximate string matching. For example,
approximate string matching has been proposed
in one of the earliest paper on music retrieval
(Ghias, Logan, Chamberlin & Smith, 1995) while
Dynamic Time Warping has been proposed in Hu
and Dannenberg (2002). Statistical approaches
have been proposed as well, in particular Markov
chains (Birmingham, Dannenberg, Wakefield,
Bartsch, Bykowski & Mazzoni, 2001) and hid-
den Markov models (Shifrin, Pardo, Meek &
Birmingham, 2002). The advantage of these ap-
proaches is that the difference between the query
and the documents can be modeled, considering
explicitly all the possible mismatches. Thus very
high performances in terms of retrieval effective-
ness can be achieved. On the other hand, all these
techniques require that the string representing the
query is matched against all the documents in
the collection, giving a complexity that is linear
with the number of documents in the collection.
Scalability to large collections of millions of
documents becomes then an issue.
For this reason alternative approaches have
been proposed that take advantage from indexing
(Doraisamy & Rüger, 2004; Downie & Nelson,
2000; Melucci & Orio, 2004; Pienimäki, 2002).
Moreover, other IR techniques can be applied to
music retrieval. For instance, Hoashi, Matsumoto
and Inoue (2003) applied relevance feedback
to a melodic retrieval task, with the main goal
of personalization of the results. The metaphor
of navigation inside a collection of documents,
which corresponds to document browsing, has
also been proposed (Blackburn & DeRoure, 1998).
On the other hand, indexing is also widely used
to retrieve or recognize music in audio format,
in particular for audio fingerprint and audio wa-
termarking techniques (Cano, Batlle, Kalker &
Haitsma, 2005).
This chapter describes some aspects of content-
based indexing, as opposed to metadata indexing,
giving a review of its basic concepts and going
in more detail about some key aspects, such as
the consistency at which candidate index terms
are perceived by listeners, the effectiveness of
alternative approaches to compute indexes, and
how individual indexing schemes can be combined
together by applying data fusion approaches.
metadata vs. content-Based
IndexIng
The first problem that arises when choosing an
indexing scheme for a music collection regards
the most effective representation of documents
content, in particular whether documents have to
be described by external metadata or directly by
a synthetic representation of their content. Both
approaches have positive and negative aspects.
Metadata usually requires extensive manual
work for retrieving external information on the
documents and for representing in a compact
way most of the subtleties of document content,
but it increases the cost of indexing and does not
guarantee consistency when different documents
are indexed by different persons. Automatic com-
putation of metadata based on external resources
has been proposed in systems for collaborative
filtering aimed, for example, at recommendation
systems, but the results are in terms of similarity
between documents and are biased by the pres-
ence of scattered data (Stenzel & Kamps, 2005).
At the state of the art they do not seem suitable
for a retrieval task. Content-based indexing is car-
ried out starting from a set of features extracted
automatically from the document itself, and it is
the main focus of this chapter.
metadata
For most media, such as images and video, the
choice of textual metadata proved to be par-
ticularly effective. Textual metadata as a tool to
describe and indexing music is a natural choice
that has been made for centuries (Dunn & Mayer,
1999). In general metadata, especially in the form
Search WWH ::
Custom Search