Staff Papers

MusicAustralia : Experiments with DC.Relation

A paper by Dr Marie-Louise Ayres, Project Manager, MusicAustralia, National Library of Australia, at the DC-ANZ Conference in Canberra, February 2003

Background

MusicAustralia is an initiative of the National Library of Australia and ScreenSound Australia, National Screen and Sound Archive 1.  The service will provide users with one-stop discovery access to Australian music resources held or managed by organisations across the library, archive, museum and commercial sectors, and across a range of digital and non-digital formats.

MusicAustralia also aims to provide authoritative directory based access to information about people, organisations and services associated with Australian music, and will interact with specialised information services, including Australia Dancing, PictureAustralia, and the Register of Australian Archives and Manuscripts 2, and with much larger services, such as the National Bibliographic Database 3.

The subject of this paper is the pilot developed in 2002, which tested the technical requirements for delivering single point discovery access to digital music objects in a range of formats - digitised printed music; digital scores, digitised audio.  In particular, the paper focuses on experiments with the DC.Relation element to support desired MusicAustralia functionality.

Pilot

The MusicAustralia pilot included 170 digital music objects from four organisations: the National Library, ScreenSound Australia, the Australian Music Centre 4 and Australian Music Online, a new music marketing initiative of the Australia Council 5.  Each of the four participants describe their digital music objects using different record formats:

  • The National Library’s records are in MARC
  • ScreenSound’s records are in the MAVIS format, which was developed specifically for audiovisual materials, and is also used by the Library of Congress
  • The Australian Music Centre’s records are in a rich non-MARC format, conforming to AACR2 Level 3 descriptive standards
  • Australian Music Online’s records are in a local XML format

To support one stop discovery, these disparate record formats were mapped to a qualified Dublin Core element set developed by a joint National Library and ScreenSound team, encoded as XML, stored in a metadata repository and indexed for searching.  The digital objects referred to in the records are delivered through local delivery systems on the custodial organisation’s server - they are not stored in a single digital object storage system. 

Those familiar with the PictureAustralia discovery model will recognise this approach.  It has some very important advantages:

  • It enhances the visibility of a range of institutional resource collections
  • It contextualises objects held across a range of collections
  • It retains organisational ‘ownership’ and branding of collection items
  • It overcomes the problems inherent in multiple data formats being used to describe similar materials
  • and its overheads for contributors are quite low.

In short, the discovery model developed for PictureAustralia and followed by the MusicAustralia pilot is cheap, effective and overcomes organisational and sectoral ‘silos’.  It is not surprising that this model has been adopted by other countries including Canada, and that it is a source of some wonder to our better-resourced colleagues in the United States, where the collaborative culture is much less developed.

While the underlying PictureAustralia discovery model is valid for MusicAustralia, the MusicAustralia pilot encompassed additional complexities.  The records contributed to MusicAustralia contained rich information that we did not want to lose. 

ScreenSound records, for instance, include important information on the roles of various agents in relation to a particular recording.  Many records include recording and release dates.  The digital objects themselves are in multiple formats, many of which are harder to deliver than single images.  These formats need to be correctly identified for a number of reasons, including the desire to provide an attractive user interface, and to generate appropriate labels - e.g. ‘Label’ rather than ‘Publisher’ for commercial sound recordings.  Many of the objects have significant and important ‘part’ relationships: printed music albums, LPs and CDs can all contain a number of separate items.  The records contributed to the pilot include subject and genre information, and standard identification numbers from a number of schemes.  The desire to include all this important and granular information led to a heavily qualified Dublin Core set - and to significant frustration at the information loss inherent in mapping rich records to a relatively flat structure.

DC.Relation

The most significant pilot complexity, however, was the desire to test the feasibility of providing support for user navigation between related versions of a music work.  Music inherently exists in both notated and audio representations.  Ideally, a national music discovery service will expose the complex relationships - between score and audio and between different score versions and different audio versions of a piece of music- which exist in the music universe.

An important sub-project of the pilot, therefore, was to test assumptions about the nature of version relationships between music items, to investigate the most suitable metadata structures to expose relationships 6, and to assess the feasibility of automating relationships through matching algorithms.  This paper reports on those experiments and their outcomes.

Methodology

The process for this part of the project consisted of three consecutive tasks:

  • Selection.  We selected primarily pre 1930 popular music across a range of subject’s areas (e.g. war, romance, patriotic, national songs, location songs).  Most of our selections were songs, but we did include non-vocal music, folklore and some manuscript and text items.  We deliberately selected content because of known or suspected version relationships, with a wide range of Waltzing Matilda materials selected as a special case study.  Selecting and digitising specific content was very resource intensive, as were record-keeping, checking and permissions processes to ensure that a significant proportion of multi-version music was included.
  • Auditing.  We audited every item selected for the content - viewed every score, listened to every performance recording, and viewed and listened simultaneously where bibliographic evidence or musicological knowledge suggested a possible relationship.  We kept detailed notes of relationships between items, including qualitative information and some subjective judgements.  The process of auditing for and recording relationships was very time-consuming - and a great pleasure.
  • Coding Relationships.  We hand-coded DC.Relation data for every relationship identified, and added this data to records previously converted from institutional data formats to the qualified Dublin Core element set agreed upon for the MusicAustralia pilot.  We used the following DC.Relation qualifiers:
  • hasPerformance/isPerformanceOf (used where there was a clear version relationship between a score and a sound recording)
  • isVersionOf (used where there was a clear version relationship between two scores)
  • isRelatedTo (used for two or more sound recordings where each sound recording had an isPerformanceOf relationship to a single score; used for clearly related works that could not be said to be versions of each other)
  • isPartOf (in this case, used only to carry MARC series data, rather than the album/segment information we actually require)
  • isFormatOf (in this case, used only to carry data alluding to preservation copies etc.)

Again, this was very time-consuming, especially as many items had multiple relationships.  This paper does not deal with the issue of ‘part relationships’: the pilot, however, highlighted existing concerns about representation of complex objects in the digital environment.  This issue requires further investigation in the context of both the MusicAustralia service itself, and for the complex object delivery systems it depends upon.

Evaluating automated matching potential.  The Australian music ‘universe’ is quite large (with more than 50 000 scores and more than 20 000 audio recordings currently recorded in the National Bibliographic Database and an unknown but presumably large number of items not recorded in the NBD).  The human resources required for manually undertaking all relationship matching and recording are simply not available in any Australian sector.  It was important, therefore to analyse the enablers and blockers for deriving desired relationships data using automated procedures.

Results

The pilot included 170 records for:

  • 102 scores
  • 65 sound recordings
  • 2 text items; and
  • 1 picture

from four organisations:

  • National Library: 111 items (score and sound)
  • ScreenSound Australia: 51 (sound only)
  • Australian Music Centre: 7 (score and sound)
  • Australia Council: 1 (sound only)

Our auditing process revealed that:

  • 36 items had no version relationships to other pilot content 7
  • 134 items had relationships with 1 or more pilot items
  • These 134 items related to 60 distinct music ‘works’, including 4 distinguishable Waltzing Matilda ‘works’.
  • These 134 items were associated with 240 identified relationships.
  • 126/240 of these relationships could have been automated by matching Creator and Title
  • 114/240 relationships could not have been automated, and would require human knowledge and intervention to establish and record relationships.

What we learned

At the outset of the pilot, we assumed we would expose a high number of ‘exact matches’ between score and performance, and that these exact matches would be candidates for straightforward ‘synchronisation’ of score and audio, using SMIL and other multimedia integration tools.  In fact, the only exact matches between score and performance were for the contemporary classical works contributed by the Australian Music Centre.  Relationships between score and performance for the predominantly popular music content varied from a few reasonably close relationships to very loose relationships indeed. 

We therefore resiled from earlier plans to present users with tight ‘packages’ of score and audio as the relationships we identified were not sufficiently close to musically or intellectually justify such packaging, and the resources required to do so were not available.  The pilot service allows users to select multiple relationships, with the option of viewing any music score whilst listening to any audio file.  The production service will follow this model, and we anticipate that organisations or cross-organisation partnerships will choose to synchronise score and audio for iconic items only, mostly for exhibition purposes.

Initially, we proposed and permitted only two Relation version qualifiers:  ‘hasPerformance/isPerformanceOf’ and ‘isVersionOf’.

Our auditing, however, revealed that the majority of identifiable pilot relationships were so loose that the relationship terminology ‘hasPerformance/isPerformanceOf’ was found to be seriously misleading.  The relationship terminology ‘isVersionOf’ to code relationships between score editions was also found to be misleading.  When we identified that scores were related, we did not mean that the scores were versions of each other, but that they were versions of an abstract but understood concept of a musical ‘work’.

We quickly realised that we needed to add a third Relation qualifier, ‘isRelatedTo’, as we had no way of coding relationships between two sound recordings that were each loose ‘performances’ of a score – these sound recordings were clearly not ‘versions’ of each other. 

We also wanted to identify non-version relationships between works.  This included very loose relationships such as recognisable snatches of melody appearing in folkloric recordings: many of these relationships could never have been identified bibliographically, but were identified melodically by an expert folklorist.

Because all relationships were horizontal, reciprocal, and ‘one to one’, the final result was a plethora of relationships, which were difficult to understand, code and test resulted.  As this series of horizontal relationships appeared to be inadequate and cumbersome, we investigated hierarchical relationship models. 

Uniform titles are employed in international cataloguing practice as a ‘binder’ for many related items.  However, the uniform title model is generally used only to aggregate individual instances (e.g. publications) of a work to the binding uniform title.  While uniform titles can be used to model some basic version information (e.g. Voss. German. for German translations of Patrick White’s novel, Voss), records employing uniform title do not generally distinguish between intellectual versions (expressions) and physical or electronic editions (manifestations). 

The FRBR (Functional Requirements for Bibliographic Records 8) information model published in 1998 models both expressions and manifestations, and is the subject of considerable work in the library and archives world. 

The FRBR model includes four entities:

  • Work (abstract but understood)
  • Expression (distinct intellectual or artistic version e.g. different words or instrumentation)
  • Manifestation (physical or online publication e.g. printed score, digital surrogate) 
  • Item (item in hand)

We undertook a short literature survey on FRBR 9 and music, to determine how others have modelled version relationships.  The FRBR/Music community is grappling with a number of modelling issues.  There is as yet no consensus on these issues, but a significant body of literature is now available, and music issues are being raised in the FRBR email list 10. Those of interest to MusicAustralia include:

  • What is the abstract Work, especially for vocal music?  Patrick Le Boeuf of the Biblioteche Nationale characterises this as the problem:
  • Primo la musica?
  • Primo le parole?
  • Primo la musica e le parole?

In the context of Waltzing Matilda, this could be expressed as a choice between modelling either:

  • Music W1 Waltzing Matilda [Composer: Christina MacPherson]
  • Music W2 Waltzing Matilda [Composer: Marie Cowan]
  • Lyric W1 Waltzing Matilda [Lyricist: A.B. Paterson]

Dave de Hugard performance is an Expression of Music W1 and of Lyric W1

Peter Dawson performance is an Expression of Music W2 and of Lyric W1

George Trevare’s Jazz Group’s performance is an Expression of of Music W2

Or:

  • W1 Waltzing Matilda [Composer: Christina MacPherson; lyricist A.B. Paterson]
  • W2 Waltzing Matilda [Composer: Marie Cowan; lyricist A.B. Paterson]
  • W3 Waltzing Matilda [Composer: Marie Cowan; no lyrics]

Dave de Hugard performance is an Expression of W1

Peter Dawson performance is an Expression of W2

George Trevare’s Jazz Group’s performance is an Expression of W3 11

In an ideal data model, music works and lyric works would be separately modelled, and a single ‘manifestation’ of a song (e.g. published score or released track) could be modelled as manifesting expressions of two distinct works, or even, following a case in the MusicAustralia pilot, a single expression of a music work with two expressions of a lyric work, I’m going back to Yarrawonga/I’m going back to Wanganui! 12  This would be particularly useful in modelling expressions of lyrics that have also been published (expressed and manifested) as poems – Waltzing Matilda is an excellent example.  The current FRBR list consensus appears to be that music and lyric works should be separately recorded where this is important to do so.  This would mean that a cataloguing record for a ‘manifestation’ would need represent the fact that the manifestation is a container for expressions and manifestations of both the music work and the lyric work.

Projects such as Levy 2 13 – which is using MCR and OCR processes to derive both MIDI files and text files from digitised printed music, and which plans to create indexed and searchable databases of both sets of coded data 14 – will presumably need to address this problem, as the project has disaggregated a piece of printed song music into its two constituent parts.  It should be noted that the ability to search for matching melodies and text elements via these databases would itself contribute to information about version relationships, which could be fed back into metadata descriptions.  Similarly, analytical processes applied to other music facets - harmony; timbre; rhythm etc - may enrich our knowledge of relationships.

Even if we could overcome this ‘two in one’ problem, a number of modelling issues remain:

  • Should all notated and performed expressions of music be modelled as a single expression category:
  • W1 Waltzing Matilda [Composer: Christina MacPherson]
  • E1 Christina MacPherson’s manuscript score
  • E2 Dave de Hugard’s performance
  • Should expressions themselves be further modelled to include sub-categories for notated and performed expressions:
  • W1 Waltzing Matilda [Composer: Christina MacPherson]
  • Notated E1 Christina MacPherson’s manuscript score
  • Performance E1 Dave de Hugard’s performance
  • Should performed expressions based on particular notated expressions be modelled as expressions of expressions:
  • W1 Waltzing Matilda [Composer: Christina MacPherson]
  • E1 Christina MacPherson’s manuscript score
  • E1 Dave de Hugard’s performance (performance is based directly on research on MacPherson’s manuscript notation)

For music in the Western art tradition, modelling performance expressions as subsequent expressions of notated expressions has considerable merit, especially for scholars trying to demonstrate relationships between particular notated editions and particular performances. 

However, in the popular, jazz and folk traditions, it is almost impossible to argue for a linear relationship between score first and performance second.  In some cases, printed music scores are clearly based on performances, i.e. the notated expression is subsequent to the performance expression.  In many cases, it is impossible to determine which came first, as dates of publication and performance are inexact, or occur in the same year.  In any case, our analysis of the MusicAustralia pilot highlighted the looseness of relationship between score and performance: popular music performances are relatively independent of notation.  Time signatures, instrumentation, words and structure are routinely changed in the space between notation and performance. 

In the MusicAustralia context, in which existing catalogue records are being re-purposed, these modelling issues are largely academic at the moment, as most of the originating records do not include sufficient information to establish any Expression level attributes, let alone the more complex models alluded to above.

We almost certainly need to decide that ‘Primo la musica’, and undertake any aggregation - at least in the first instance - to our shared understanding of an identifiable music Work, which is bibliographically identifiable in many instances.  The Variations2 15 project at the University of Indiana has adopted this approach, modelling Works and Instantiations, but not - for the moment - modelling Expressions.  It is likely that international theory and practice will decide that the most simples and most sustainable approach to Expressions is to model all versions as expressions of equal ‘weight’, but to allow specific annotated links between expressions of a single work where the relationship is sufficiently strong.  We will be closely monitoring international practice in this regard.

Automating Relationships

As MusicAustralia does not incorporate processes for music or textual transcription and analysis (as, for example, the Levy 2 project does), any automated matching of items and aggregation to Expression and Work levels can only occur at the bibliographic level. 

Barriers to automating aggregation included:

  • Mismatching title and author information.  A few of these mismatches were cataloguer error; most arose from different sectors using different (or no) name authority files, and from the fact that music titles can change sufficiently to prevent automated matching.
  • Lack of granularity in bibliographic records for publication containers. Desired relationships are between items that are parts of publication containers such as albums and CDs.  In many MARC records, individual titles are included only in contents notes rather than as added title or name/title entries.  For example, we would be unable to create an explicit relationship between the printed score of the song Banish the Budget Blues and its commissioned performance by the Song Company, unless we attempted to match title fields with text in the contents notes of the National Library MARC records for each item.  An additional concern here is that content titles are frequently abbreviated and sometimes contain extraneous text such as numbers that prevent matching.
  • Lack of granularity in bibliographic records for unpublished material containers. The desired relationships are between items that are parts of non-publication containers such as manuscript collections and oral history recordings.  Titles of individual music works are often ‘buried’ within a finding aid or a transcript. 
  • Desired relationships are between related works, rather than between expressions of works.  For example, we would be unable to automate a relationship between any Waltzing Matilda works and the Geoff Pryor cartoon ‘Once a Jolly Swagman’, as there is no reference to Waltzing Matilda, Christina MacPherson, Marie Cowan or A.B. Paterson in the Pryor record.  Similarly, the melodic relationships identified by Dave de Hugard between 19th century printed music and 20th century folkloric performances require expert knowledge, as there are no bibliographic clues that these items - or excerpts of items - might be related.

We concluded that the potential to automate aggregation of manifestation level records could increase if:

  • descriptive metadata for unpublished ‘collections’ improves, e.g. if EAD finding aids for archival collections are available, include title coding, and can be searched.
  • individual contents of publication containers such as albums and multiple item sound recordings were fully described with name and title entries, and can stand as records in their own right.
  • more shared name and date authority work is undertaken across institutions and sectors.

We also concluded that in order to expose these Work, Expression and Manifestation entities in MusicAustralia displays:

  • there will be a residual need for a workflow process and interface to allow human inspection and aggregation of items to Expression and Work levels.
  • There will be a residual need for a workflow process and interface to allow such human knowledge to be recorded in the form of expression annotations and explicit linking of very strongly related expressions.

These lessons will be incorporated into the statement of requirements for the MusicAustralia discovery service, and those lessons that relate to contributors’ record quality and granularity, have already informed recommended metadata standards for MusicAustralia contributors.

Conclusions

Dublin Core was chosen for the MusicAustralia pilot because it has proven utility as an interchange format for national discovery services that include contributor records in a range of formats. 

The MusicAustralia pilot, however, demonstrated that Dublin Core is not the most suitable interchange format for this service for two key reasons:

  • Unlike the PictureAustralia example, a large number of MusicAustralia contributor records are likely to be rich, high quality, granular and in well-recognised data formats such as MARC and MAVIS.  ‘Dumbing’ these records down to Dublin Core results in unacceptable losses of rich music information.
  • The DC.Description.Version qualification can be used to identify different editions etc. of a work using text.  The DC.Relation element can be used to display and support navigation between items with flat, horizontal relationships.  However, the kinds of relationships MusicAustralia wants to expose are a combination of vertical and horizontal relationships, and rely heavily on abstract but well understood and demonstrable concepts of the Work and the Expression or version.  At this stage, DC does not offer support exposure of navigational pathways that explicitly acknowledge both vertical and horizontal relationships.

This does not mean that MusicAustralia and DC will be strangers in the future.  We anticipate that some organisations - especially museums - will describe their music objects using DC, and that others - especially the commercial sector - will describe their music objects using local XML data formats.  But we already know that the majority of contributed records will be in rich formats such as MARC and MAVIS.  We therefore need to find a richer interchange format and are investigating MODS for this purpose.  As a MODS to MARC conversion tool already exists, it would be possible to convert MODS records to MARC, thus supporting inclusion of a much wider set of originating data formats in the National Bibliographic Database, as well as in MusicAustralia.  We also anticipate exposing MusicAustralia data to other information services that may not have the same for rich data and relationship requirements, so MusicAustralia records may in the future be made available in DC formats and other formats as required.

We are also investigating whether existing MARC/MODS fields can be used to produce ‘good enough’ FRBR for the medium term.  This includes investigation of using Uniform Title or linking fields to expose vertical Work and Expression relationships, and of populating linking fields with data to expose relationships between works. 

Given the inevitable human work required to enrich records to support this functionality, this experimentation depends largely on whether MusicAustralia contributors believe it will be useful for users, and on their ability to resource that belief. 

Footnotes

  1. http://www.nla.gov.au and http://www.screensound.gov.au respectively

  2. http://www.australiadancing.org, http://pictureaustralia.org, and http://www.nla.gov.au/raam/ respectively.

  3. Accessible via the Kinetica service:  http://www.nla.gov.au/kinetica/

  4. The Australian Music Centre represents more than 400 contemporary Australian composers and collects and disseminates their music in score and audio formats.  See: http://www.amcoz.com.au

  5. Australian Music Online is available at: http://www.amo.org.au.  AMO is an initiative under the Australia Council’s Audience and Market Development program.  See: http://www.ozco.gov.au/audience/index.htm

  6. A proposal to qualify DC.Description with ‘Version’ was published by the DC Libraries Working Group in April 2002, and qualification now has the ‘Recommended’ status.  The aim of this proposal was to allow inclusion of specific edition information, such as that in the MARC 250 field.  The proposal demonstrates awareness of FRBR entities in its comment :’ Includes statements designating an edition or version. This may describe the resource in terms of a number or statement that assists in distinguishing it from other expressions of similar content’, and briefly alludes to use of the DC.Relation element : ‘Relation, except that relation is a pointer to some other resource and is not a descriptor of its version. There is also not necessarily any related item to link to’, but does not explicitly mention strategies for relating editions to abstract expression and work entities – in other words, what the relationship between DC.Description.Version and DC.Relation might be.  The proposal is available at: http://www.loc.gov/marc/dc/description-version_prop.html

  7. Previous investigations of large music databases suggest that as many as 8/10 music works exist in more than one version.  Reference required.

  8. Final report : www.ifla.org/VII/s13/frbr/frbr.pdf

  9. See IFLA’s online FRBR Bibliography: http://www.ifla.org//VII/s13/wgfrbr/bibliography.rtf

  10.  Moderated by Patrick le Boeuf.  List address is: FRBR@bnf.fr

  11. All available at the MusicAustralia pilot site: http://www.musicaustralia.org

  12. McBeath, Neil, ‘I’m going back to Wanganui; I’m going back to Yarrawonga’, Sydney: J.Albert & Son, 1919.  Available online at:  http://nla.gov.au/nla.mus-an6052667

  13. http://www.dml.indiana.edu/

  14. See S. Choudhury et.al, ‘Strike up the score: deriving searchable and playable digital formats from sheet music, DLib Magazine, February 2001, Vol 7, No 2, http://www.dlib.org/dlib/february01/choudhury/02choudhury.html

  15.  See: http://www.dml.indiana.edu/

Back to the top