DBTune blog

Last.fm recently unveiled their new fingerprinting lookup mechanism. They did aggregate quite a lot of fingerprints (650 million) using their fingerprinting software, which is a nice basis for such a lookup, perhaps bringing a viable alternative to Music DNS. I gave it a try (I just had to build a Linux 64 version of the lookup software), and was quite surprised by the results. The quality of the fingerprinting looks indeed good, but here are the results for a particular song:

<?xml version="1.0"?>
<!DOCTYPE metadata SYSTEM "http://fingerprints.last.fm/xml/metadata.dtd">
<metadata fid="281948" lastmodified="1205776219">
<track confidence="0.622890">
    <artist>Leftover Crack</artist>
    <title>Operation: M.O.V.E.</title>
    <url>http://www.last.fm/music/Leftover+Crack/_/Operation%3A+M.O.V.E.</url>
</track>
<track confidence="0.327927">
    <artist>Left&ouml;ver Crack</artist>
    <title>Operation: M.O.V.E.</title>
    <url>http://www.last.fm/music/Left%C3%B6ver+Crack/_/Operation%3A+M.O.V.E.</url>
</track>
<track confidence="0.007860">
    <artist>Leftover Crack</artist>
    <title>Operation MOVE</title>
    <url>http://www.last.fm/music/Leftover+Crack/_/Operation+MOVE</url>
</track>
<track confidence="0.006180">
    <artist>Leftover Crack</artist>
    <title>Operation M.O.V.E.</title>
    <url>http://www.last.fm/music/Leftover+Crack/_/Operation+M.O.V.E.</url>
</track>
<track confidence="0.004883">
    <artist>Leftover Crack</artist>
    <title>Operation; M.O.V.E.</title>
    <url>http://www.last.fm/music/Leftover+Crack/_/Operation%3B+M.O.V.E.</url>
</track>
<track confidence="0.004826">
    <artist>Left&ouml;ver Crack</artist>
    <title>Operation M.O.V.E.</title>
    <url>http://www.last.fm/music/Left%C3%B6ver+Crack/_/Operation+M.O.V.E.</url>
</track>
<track confidence="0.004717">
    <artist>Left&ouml;ver Crack</artist>
    <title>13 - operation m.o.v.e</title>
    <url>http://www.last.fm/music/Left%C3%B6ver+Crack/_/13+-+operation+m.o.v.e</url>
</track>
....
</metadata>

And it goes on and on... There are 21 results for this single track, which all actually correspond to this track.

So, what is disturbing me here? After all, the first result holds textual metadata that I could consider as somehow correct (even if that's not the way I would spell this band's name, but they plan to put a voting system to solve this sort of issues).

The real problem is that there are 21 URI in last.fm for the same thing. The emphasis of the last.fm metadata system is then probably on the textual metadata: two different ways of spelling the name of a band = two bands. But I do think it is wrong: for example, how would you handle the fact that the Russian band Ария is spelled Aria in English? The two spellings are correct, and they correspond to one unique band.

In my opinion, the important thing is the identifier. As long as you have one identifier for one single thing (an artist, an album, a track), you're saved. The relationship between a band, an artist, a track, etc. and its label is clearly a one-to-many one: the quest for a canonical spelling will never end... And what worries me even more is that it tends to kill the spellings in all languages but English (especially if a voting system is in place?).

Once you have a single identifier for a single thing within your system, you can start attaching labels to it, perhaps with a language tag. Then, it is up to the presentation layer to show you the label matching your preferences. And if you tend for such a model, Musicbrainz (centralised and moderated) or RDF and the Music Ontology (decentralised and not moderated) are probably the way to go.

I guess this emphasis on textual metadata is mainly due to the ID3 legacy and other embedded metadata format, which allowed just one single title for the track, the album and the artist to be associated with an audio-file?

I think that the real problem for last.fm will now be to match all the different identifiers they have for a single thing in their system, which is known as the record linkage problem in the database/Semantic Web community. But I also think this is not too far-fetched, as they already began to link their database to the Musicbrainz one?

Chris Sutton and I did some work since the first release of GNAT, and it is now in a releasable state!

You can get it here.

What does it do?

As mentioned in my previous blog post, GNAT is a small software able to link your personal music collection to the Semantic Web. It will find dereferencable identifiers available somewhere on the web for tracks in your collection. Basically, GNAT crawls through your collection, and try by several means to find the corresponding Musicbrainz identifier, which is then used to find the corresponding dereferencable URI in Zitgist. Then, RDF/XML files are put in the corresponding folder:

/music
/music/Artist1
/music/Artist1/AlbumA/info_metadata.rdf
/music/Artist1/AlbumA/info_fingerprint.rdf
/music/Artist1/AlbumB/info_metadata.rdf
/music/Artist1/AlbumB/info_fingerprint.rdf

What next?

These files hold a number of <http://zitgist.com/music/track/...> mo:available_as <local file> statements. These files can then be used by a tool such as GNARQL (which will be properly released next week), which swallows them, exposes a SPARQL end point, and provides some linked data crawling facilities (to gather more information about the artists in our collection, for example), therefore allowing to use the links pictured here (yes, sorry, I didn't know how to introduce properly the new linking-open-data schema - it looks good! and keeps on growing!:-) ):

Two identification strategies

GNAT can use two different identification strategies:

• Metadata lookup: in this mode, only available tags are used to identify the track. We chose an identification algorithm which is slower (although if you try to identify, for example, a collection with lots of releases, you won't notice it too much, as only the first track of a release will be slower to identify), but works a bit better than Picard' metadata lookup. Basically, the algorithm used is the same as the one I used to link the Jamendo dataset to the Musicbrainz one.
• Fingerprinting: in this mode, the Music IP fingerprinting client is used in order to find a PUID for the track, which is then used to get back to the Musicbrainz ID. This mode is obviously better when the tags are crap :-)
• The two strategies can be run in parallel, and most of the times, the best identification results are obtained when combining the two...

Usage

• To perform a metadata lookup for the music collection available at /music:

./AudioCollection.py metadata /music

• To perform a fingerprint-based lookup for the music collection available at /music:

./AudioCollection.py fingerprint /music

• To clean every previously performed identifications:

./AudioCollection.py clean /music

Dependencies

• MOPY (included) - Music Ontology PYthon interface
• genpuid (included) - MusicIP fingerprinting client
• rdflib - easy_install rdflib
• mutagen - easy_install mutagen
• Musicbrainz2 - You need a version later than 02.08.2007 (sorry)

This post deals with further interlinking experiences based on the Jamendo dataset, in particular equivalence mining - that is, stating that a resource in the Jamendo dataset is the same as a resource in the Musicbrainz dataset.

For example, we want to derive automatically that http://dbtune.org/jamendo/artist/5 is the same as http://musicbrainz.org/artist/0781a... (I will use this example throughout this post, as it illustrates many of the problems I had to overcome).

Independent artists and the failure of literal lookup

In my previous post, I detailed a linking example which was basically a literal lookup, to get back from a particular string (such as Paris, France) to an URI identifying this geographical location, through a web-service (in this case, the Geonames one). This relies on the hypothesis that one literal can be associated to exactly one URI. For example, if the string is just Paris, the linking process fails: should we link to an URI identifying Paris, Texas or Paris,France?

For mainstream artists, having at most one URI in the Musicbrainz dataset associated to a given string seems like a fair assumption. There is no way I could start a band called Metallica, I think :-)

But, for independent artist, this is not true... For example, the French band Both has exactly the same name as a US band. We therefore need a disambiguation process here.

Another problem arises when a band in the Jamendo dataset, like NoU, is not in the Musicbrainz dataset, but there is another band called Nou there. We need to throw away such wrong matchings.

Disambiguation and propagation

Now, let's try to identify whether http://dbtune.org/jamendo/artist/5 is equivalent to http://zitgist.com/music/artist/078... or http://zitgist.com/music/artist/5f9..., and that http://dbtune.org/jamendo/artist/10... is not equivalent to http://zitgist.com/music/artist/7c4....

By GETting these URIs, we can access their RDF description, which are designed according to the Music Ontology. We can use these descriptions in order to express that, if two artists have produced records with similar names, they are more likely to be the same. This also implies that the matched records are likely to be the same. So, at the same time, we disambiguate and we propagate the equivalence relationships.

Algorithm

This leads us to the following equivalence mining algorithm. We define a predicate similar_to(+Resource1,?Resource2,-Confidence), which captures the notion of similarity between two objects. In our Jamendo/Musicbrainz mapping example, we define this predicate as follows (we use a Prolog-like notation---variables start with an upper case characters, the mode is given in the head: ?=in or out, +=in, -=out):

     similar_to(+Resource1, -Resource2, -Confidence) is true if
               Resource1 is a mo:MusicArtist
               Resource1 has a name Name
               The musicbrainz web service, when queried with Name, returns ID associated with Confidence
               Resource2 is the concatenation of 'http://zitgist.com/music/artist/' and ID

and

     similar_to(+Resource1, +Resource2, -Confidence) is true if
               Resource1 is a mo:Record or a mo:Track
               Resource2 is a mo:Record or a mo:Track
               Resource1 and Resource2 have a similar title, with a confidence Confidence

Moreover, in the other cases, similar_to is always true, but the confidence is then 0.

Now, we define a path (a set of predicates), which will be used to propagate the equivalence. In our example, it is {foaf:made,mo:has_track}: we are starting from a MusicArtist resource, which made some records, and these records have tracks.

The equivalence mining algorithm is defined as follows. We first run the process depicted here:

Every newly appearing resource is dereferenced, so the algorithm works in a linked data environment. It just uses one start URI as an input.

Then, we define a mapping as a set of tuples {Uri1,Uri2}, associated with a confidence C, which is the sum of the confidences associated to every tuple. The result mapping is the one with the highest confidence (and higher than a threshold in order to drop wrong matchings, such as the one mentioned earlier, for NoU).

Implementation

I wrote an implementation of such an algorithm, using SWI-Prolog (everything is GPL'd). In order to make it run, you need the CVS version of SWI, compiled with the http, the semweb and the nlp packages. You can test it by loading ldmapper.pl in SWI, and then, run:

?- mapping('http://dbtune.org/jamendo/artist/5',Mapping).

To adapt it to other datasets, you just have to add some similar_to clauses, and define which path you want to follow. Or, if you are not a Prolog geek, just give me a list of URI you want to map, along with a path, and the sort of similarity you want to introduce: I'll be happy to do it!

Results

I experimented with this implementation, in order to automatically link together the Jamendo and the Musicbrainz dataset. As the current implementation is not multi-threaded (it runs the algorithm on one artist after another), it is a bit slow (one day to link the entire dataset). It derived 1702 equivalence statements (these statements are available there), distributed over tracks, artists and records, and it spotted with a good confidence that every other artist/track/record in Jamendo are not referenced within Musicbrainz.

Here are some examples:

• http://dbtune.org/jamendo/artist/5 (here, the different members comes from the Musicbrainz data set, whereas the bittorrent links come from the Jamendo dataset)
• http://dbtune.org/jamendo/artist/2683
• http://dbtune.org/jamendo/record/3365 (here, the tags come from the Jamendo dataset)
• http://dbtune.org/jamendo/track/22663