Musical data can be analysed, combined, transformed and exploited for diverse purposes. However, despite the proliferation of digital libraries and repositories for music, infrastructures and tools, such uses of musical data remain scarce. As an initial step to help ll this gap, we present a survey of the landscape of musical data on the Web, available as a Linked Open Dataset: the musoW dataset of catalogued musical resources. We present the dataset and the methodology and criteria for its creation and assessment. We map the identi ed dimensions and parameters to existing Linked Data vocabularies, present insights gained from SPARQL queries, and identify signi cant relations between resource features. We present a thematic analysis of the original research questions associated with surveyed resources and identify the extent to which the collected resources are Linked Data-ready.
Since the early stages of its development, the Web has o ered opportunities as a
platform to disseminate and exchange information for research and scholarship
in the humanities. The digitisation of physical archives, records and other
artefacts relevant to humanities research has enabled novel approaches and methods
of enquiry that involve computation as a core component, acting on digitized
collections of texts, numbers, images, and diagrams [
The primary research question is: what is the status of musical data with respect to the Web of Data? Secondarily: to what extent are musical resources ready to be published and linked on the LOD cloud? What types of research and enquiry are musical data meant for, and what direction Semantic Web research should take in order to support them (better)? Through the production of a Linked Open Dataset of musical resources published on the Web, described according to a set of key dimensions, we derive a classi cation of the available data, its nature, form and purpose, and an identi cation of distinguishing features of the di erent types of resources. In the light of the gaps of the current landscape with relation to the Web of Data, we identify a set of representative themes in musical research, and formulate hypotheses on how the Semantic Web can help with answering them. Thus we intend to contribute by inspiring possible future directions in Semantic Web developments for the humanities.
We contribute (a) a LOD dataset of catalogued musical resources, as well as a related list of signi cant SPARQL queries; (b) An analysis of the distinguishing features of each type of data; (c) a set of research themes that are the focus of data oriented musical research, extracted from the corpus; (d) and assessment of the LD-readiness of the resources, by classifying the resources with respect to the 5-Star Web of Data schema. 2
In this section we describe existing work addressing the collection of relevant and reusable musical data; the role of musical datasets in interoperable and reusable work ows in Music Information Retrieval (MIR); and the dimensions considered for analysis in existing surveys in the MIR and the Semantic Web communities.
One of the most reused datasets in MIR is the Million Song Dataset [
Musical data has a key role in the reusability and interoperability of work ows
in MIR. Page et al. [
In the Semantic Web, dataset descriptions typically deal with only one dataset,
and hence domain ontology catalogues and surveys are more relevant to the
identi cation of suitable dimensions. In [
Our survey addresses the shortcomings in existing musical data collections, by contributing a more abstract gathering of musical resources available on the Web, and thus with high reusability and a broader scope; in work ows in MIR, since we provide a way to nd and reuse those resources by means of HTTP dereferencing, as well as a way to facilitate repurposing by specifying what the original purpose of each of these resources was; and in Semantic Web dataset surveys, by borrowing analysis dimensions from at least three sources: the Semantic Web generally, ontologies from humanities research, and terms from MIR research generally. 3
In this paper, we assess the status of musical data in the Web of Data, and discuss the potential contribution of the Semantic Web to support music research. To promote reliability, we focus in the rst instance on sources derived from musical research and scholarship. The work ow of our assessment methodology is as follows: 1. We design a set of ad hoc dimensions to describe the resources of the domain, and we use these dimensions to describe the resources in a table; 2. We survey the following musical resources: repositories, digital libraries, datasets, catalogues, projects, digital editions, services, software, formats, schemas and ontologies; 3. We map dimensions and parameters to well-known Linked Data vocabularies; and we produce musoW, a Linked Data dataset describing all these resources (Section 4); 4. We query this dataset in SPARQL to draw an overview of the collection and gain insights (Section 4); 5. We conduct a statistical analysis and identify signi cant relations between resources features (Section 5); 6. We conduct a thematic analysis of the research questions associated with the above resources (Section 6); 7. We analyse the results in the light of the ve-star Linked Data principles using Formal Concept Analysis (Section 7).
In the remainder of this Section we describe the creation of the musoW dataset and the criteria for its analysis.
The survey is designed with the perspective of potential applications of musical data in the Semantic Web, and its targeted users are researchers. To gather our collection, we relied on resources created and used by researchers, and retrievable using online aggregators - which also target researchers. We look for research objects already evaluated in musicology, ensuring their reliability, and establishing a reproducible gathering criterion. We scraped these online aggregators to retrieve names of projects, URLs and descriptions. We only extract resources providing digitizations or transcriptions of scores, performance audio, and, eventually, a critical apparatus of notated music. We excluded materials for theoretical studies - such as literature, archives and libraries of resources not available online - and collections of learning materials - e.g. audio and video courses.
To implement the survey we designed a set of 46 dimensions to describe these objects, and we created a table in which such dimensions are columns whose values are validated by controlled vocabularies. A subset of dimensions is applicable to all the types of resources: resource ID, URL, description, project afliation, search criterion, resource type, reused resources (or connection to other projects), purpose (learning or research), access restrictions, licenses, situation or task, and target audience. Another subset applies to data collections only (repositories, digital libraries, datasets, catalogues, and digital editions), and includes: an item example, gathering criteria of collections (genre, artist, temporal or geographical), subject terms from both Music Ontology2 and a local controlled vocabulary, a list of services o ered by the resource (data dump, browsable interface, queryable interface, API, SPARQL endpoint), collection size, data size, which features of symbolic notation (melody, harmony, rhythm, timbre, contour or shape, structure of a song, descriptive metadata) are provided as structured data (if applicable), formats and their interoperability. Most of these dimensions are shared also with schemas, ontologies, services, software and formats, except the ones for describing the scope of contents.
Part of the purpose of the survey is to better understand the existing aims of music researchers. Consequently, as well as describing and linking existing research data, we cached the research questions associated with each surveyed resource, where explicitly available or easily inferred from project documentation on the web.3
In order to assess the extent to which musical data conforms to the Web of Data principles, characterize the landscape of musical data, and make emerge potential gaps and opportunities for further research, we chose to observe the corpus under four perspectives:
(1) Quantitative. The role of a quantitative analysis is of illustrating the musoW dataset in numbers, by aggregating items with respect to the di erent dimensions, therefore giving a picture of the musical data landscape. We observed dimensions and formulated a set of questions related to them. We implemented those as SPARQL queries, we report the major ndings in Section 4.
(2) Statistical. We performed statistical analysis in order to understand some of the relationships between the dimensions. Our analysis focussed on answering questions related to the size and resource types of the collections and how that related to their scope and musical features.
(3) Thematic. We performed a thematic analysis of the research questions associated with the surveyed resources. This involved coding the statements contained in the research questions and then clustering the codes into a series of emerging themes related to music research.
(4) LD-Readiness. We analysed the data to assess to what extent the
collected resources are LD-ready. The 5-Stars Open Data development scheme
identi es ve key dimensions of open data4: Open Licence (OL), Machine
readable (RE), Open format (OF), Adoption of URIs (URI), and Linked Data (LD).
2 http://musicontology.com
3 When research questions could not be evidenced directly, we provided keywords
summarizing our best understanding of the purpose, task or situation.
4 5-Star Open Data: http://5stardata.info/en/
We therefore generated these ve derived dimensions from the collected data.
We analysed the resulting data using Formal Concept Analysis (FCA) with the
Contento tool [
The musoW dataset is available online5 and the content can be queried in SPARQL through the data.open.ac.uk endpoint.6 To make this analysis reproducible, we publish the SPARQL queries on which it is based. Query identi ers are reported below in squared brackets. To facilitate access to the results of these queries by any application, we also publish an equivalent RESTful API.7 The collection includes 351 resources: 187 repositories and digital libraries, 44 catalogues and 3 projects, 36 datasets, 21 digital editions, 22 software, 14 services, 12 ontologies, 2 schemas, and 3 formats.
Repositories and digital libraries are the most representative resources collecting musical data. They mainly o er digitisations of scores and lyrics (77%) [DR6], published as PDF (62%) and/or JPG (40%) [DR4]. Audio records are provided by 29% of repositories [DR7], as MIDI les (45%) and/or MP3 (29%) [DR4]. Only 20% of repositories o er structured data on symbolic music notation [DR8], representing melody (95%), rhythm (76%), harmony (74%), structure of a song (46%), timbre or contour (less than 10%) [DR14]. The most used formats are here MusicXML (46%), custum XML (23%) and MEI/XML (10%). The more the scale of repositories increases, the less structured formats for representing symbolic notation seem to be used [DR4] and the less depth of analysis is provided [DR13].8 We mainly found items belonging to the same musical genre (64%), and/or to the same country (39%) and/or falling within the same period (31%) National projects seem to a ord dealing with large amounts of data, while smaller projects narrow the scope to a single genre. Datasets are the second most represented category of resources, mainly available under Creative Commons licenses and in several interoperable formats, such as RDF (44%), JSON (14%), TXT (11%), XML and CSV (less than 10%), and others [DS2]. The scope is heterogeneous, and doesn't provide any insight on a particular or shared interest [DS4]. Instead, purposes and tasks seem to be the gathering criterion: mainly focusing on research goals (89%) [DS10], the aim is to make improvements in music analysis (28%), music information retrieval (22%) - including more speci c tasks like genre recognition, score-audio linking, and machine learning. Few ones are targeted to disciplines like musicology and history of music (11%) [DS15]. Among the RDF datasets, the focus on descriptive metadata of music is predominant (75%), while 19% represent features 5 musoW: https://github.com/enridaga/musow. 6 Named Graph: http://data.open.ac.uk/context/musow. 7 API: http://grlc.io/api/albertmeronyo/mudow-queries 8 Moreover, several large-scale repositories o er queryable structured data only for incipits of works, even though we included them in the category of such kind of data providers. extracted from audio data, and only one deals with features extracted from notated music [LD1]. To explain this, we look at tasks motivating the realisation of such datasets, nding that there is a common need of publishing a speci c kind of data otherwise not available in other data sources (44%) - e.g repositories generally do not o er a data dump - and aggregate it with information from similar datasets (25%), e.g. to enable research in domains like history of music and musicology (25%). Furthermore, music analysis (19%) and music information retrieval (12,5%) seem to nd in LOD a testing bench [LD3]; but only one dataset is reused by other music related projects [DS12]. Finally, data dumps are the most common way to publish data (75%), while only 37,5% o er a SPARQL endpoint [LD2]. Digital editions generally o er small-scale collections of musical data [DE4]. They mainly deal with scores of a single artist (76%) or contemporary related groups of artists (less than 33%) [DE3]. Less than 38% o er structured data on symbolic notated music [DE7]. Still, the most used formats are JPG (47%), PDF (33%), MEI/XML and MP3 (23%) [DE2]. The main goal of such resources is to give a contribution in elds like musicology (86%) and history of music (76%). A shared concern regards visualization of complex information like variants and genetic of music. Few tools have been developed in order to support tasks like annotation and visualisation. MEI/XML les are generally the preferred input [DE11]. Services and software are here mainly considered because of their task and possibilities of reuse. We mainly found tools for annotating music (25%), and enabling further analysis in research elds like musicology (25%), music information retrieval, history of music (19%) and music philology related issues (11%) - e.g. Optical Music Recognition, music style analysis, measure annotation. Secondly, as already revealed when describing digital editions, data visualisation is a shared concern (14%) [SS3]. 75% of such tools deal with a structured representation of music features, such as melody and rhythm (100%), harmony (96%)[SS4]. 64% of software/services extract music features directly from notated music, rather than audio tracks (33%) [SS5]. Ontologies and schemas do not o er insights on a shared need in knowledge representation at this stage of the analysis. Indeed, 36% deal with the representation of features extracted from audio tracks, 36% from descriptive metadata (e.g. cataloguing information of songs, artists, genres), and 21% from symbolic notation [SO2]. There are no evidences of a clear and shared approach to represent music knowledge extracted from audio/scores. In fact, none of the proposed models are reused in other projects than the one where they were born in9 [SO4]. 5
In order to understand some of the relationships between the dimensions of the survey, a statistical analysis was conducted, focussed on answering the following questions: 1) Is there a relationship between the size of the collection and the 9 Except the Music Ontology, which does not represent any features of notated music or audio les. types of resources it holds? 2) Are there any relationships between the musical features represented (e.g. lyrics, rhythm) and size and resource type of the collection? 3) Are there any relationships between the de ned scope of the collection (e.g. by time period, artist, genre, geography) and its size and resource type?
The relationship between size and type was analysed. To ensure su cient cell sizes in the analysis, collection size categories were merged (<100, <1000, >1000) and analysis was restricted to the four most prominent resource types (catalogues, digital libraries, digital editions and repositories). A signi cant interaction was found between size and type (Fisher exact test, p < 0.01). Essentially, digital editions tend to form smaller collections than the others. Multinomial regression analysis was used to test if musical features could predict resource type. Due to cell sizes, this was restricted to the following features: melody, rhythm, lyrics and structure. Lyrics are more likely to be found in software (B = 2.183, p < 0.01) and datasets (B = 1.448, p < 0.05) but less likely in digital libraries (B = 2.125, p < 0.05). Rhythm is more likely to be represented in digital editions (B = 2.823, p < 0.05), software (B = 4.530, p < 0.01) and datasets (B = 3.040, p < 0.01). Structure is more likely to be represented in software (B = 1.680, p < 0.05) and datasets (B = 1.711, p < 0.01). Ordinal regression analysis was used to test if musical features could predict collection size. Larger collections are more likely to feature melody (Wald = 4.178, p < 0.05). Multinomial regression analysis was also used to test if the de ned scope (e.g. by genre or artist) could predict resource type. Digital editions are more likely to be scoped in terms of artist (B = 2.655, p < 0.01). Software (B = 2.810, p < 0.01) and datasets (B = 1.022, p < 0.05) are less likely to be scoped in terms of genre. Ordinal regression analysis was used to test if scope could predict collection size. Smaller collections are more likely to be scoped in terms of artist (Wald = 28.359, p < 0.01) or genre (Wald = 7.362, p < 0.01) than larger collections.
We can see overall that: (1) there is a relationship between resource type and size; (2) musical features are more or less likely to be represented in a collection depending on its size and resource type; (3) there are relationships between the scope of the collection and its size and resource type. 6
For 37 of the projects a textual description of the research question or questions
to be answered using the dataset was identi ed. In order to characterise the range
of issues raised in the research questions, a thematic analysis [
Of particular interest is the types of musicological inquiry identi ed from the projects. These are: nding out what a class of objects (such as blues songs) have in common; understanding changes in music over time (e.g. the time the piece was written or the biblical period the piece is about); analysis of di erent versions or editions of the same piece and how they vary; analysis of heterogeneous resources associated with the same theme (e.g. documents and data about jazz artists); comparing how people work with digital versus analogue artefacts; and contrasting classes of work (e.g. Chopin versus others).
Projects aimed to develop support for di erent forms of activity such as research, teaching and performance. Research aims were concerned with supporting di erent types of music content publishing such as rendering visual scores from some underlying machine readable format and indexing scores according to this format. Research also aimed to publish musical artefacts (such as scores) with some form of associated scholarly interpretation. Some projects had a research goal to construct an archive, but of di erent types of material such as scores, recordings, ephemera and libretti. 7
We now report on the evaluation of the collected resources with respect to the 5 Star Open Data paradigm. The 5 Star Open Data scheme includes ve level of compliance with the Web of Data. To map the musoW catalogue with this scheme we generated ve derived dimensions with the following criteria: OL Open Licence. The resource is publicly accessible with an open access licence (e.g. CC-BY, CC0, OGL), also if only for human consumption.
RE Machine Readable. The resource contains structured data published in a machine readable format (although it can be a proprietary one). Resources being published in any interoperable format or through Web APIs are considerable to be machine readable.
OF Open Format. The resource is published in an open standard (e.g. CSV).10
URI URIs. The resource makes use of Uniform Resource Identi ers (URIs) to
identify the described entities. We derived these dimensions for all the
resources expressed in RDF or related vocabularies (OWL, SKOS).
LD Linked Data. The resource is published in RDF using a SPARQL endpoint.11
We built a FCA formal context including the catalogue items and the ve derived
dimensions: OL, RE, OF, URI, LD. Following the FCA approach, we generated
a concept lattice and labelled the concepts from one to ve stars using the
Contento tool [
300 200 100
0 (a) The FCA annotated lattice developed for the LDreadiness analysis 1* RE 2* OF -OL -RE 3* -OL 4* -OL 5* (b) Distribution of resources with respect to the 5* scheme. 8
Although some resources are ready to be linked to the Web of Data, the majority of resources are left behind (see Figure 1b). The lack of an open license associated with the data or collection seems to be a generalized issue, a nontechnical limitation that nevertheless hinders the reuse-ability of the resources. We notice that the more the scale of repositories increases, the less structured formats for representing symbolic notation seem to be used. This emerged especially in data sources coming from National projects, that can a ord dealing with large amounts of resources, and it might be appointed to a heterogeneity of resources' typologies. Dataset are focused on specialized research tasks (e.g. in the context of MIR), but most of them include metadata rather then musical content expressed symbolically. Although software and services for semantic lifting of musical content exists, they are not applied to large repositories or reused outside the original context, often part of small sized digital editions. These observations suggest the need of a reusable and scalable work ow to support the life cycle of musical data on the Web. More importantly, there is a lack of understanding about what kind of life cycle musical data could have on the Web, and whether it would be possible to support it with systematic approaches.
Observing digital editions, we considered that tool support for annotation, exploration and visualization of musical corpora it's still at its infancy, and we argue that Semantic technologies can have a role in the way musical content can be abstracted and organized for browsing and exploration.
We also notice the opportunity of Linked Data within music on two issues of authority: one related to notes the experts know are wrong; and the other where experts disagree, e.g. because of lack of original score, or poor handwritten originals. This shows an overlap with trust from Web data in general, for which some Linked Data approaches could be of use. In particular, the description and publication of provenance of musical research objects using the PROV vocabulary12, and the sharing of annotations on top of musical resources using the Open Annotations Data Model13, could ll the gaps in these issues.
In the light of the scarcity of Linked Data resources, at least concerning the
publication of music notation as Linked Data, for which we could only nd one
resource, there is a long way to go with respect to the reusable, repurposable
and interoperable work ows that have been proposed in MIR [
In this paper, we surveyed the landscape of musical data on the Web and presented the musoW dataset, a Linked Open Data catalogue of musical resources published on the Web with the purpose of supporting musical research and scholarship. We observed that a large amount of resources are not ready to be part of the Web of Data, and the main obstacles are due to the heterogeneity of large collections, the uncertainty in licensing, and the lack of large scale approaches to semantic lifting of musical resources and data publishing. Ultimately, it is relevant to notice a cultural bias in the distribution of how musical features are represented. In fact, larger collections are more likely to feature melody, re ecting clearly a Western-centric point of view. As all Web material, we can observe this will have issues with representative sampling and quality that could be interesting to investigate further. Furthermore, thanks to the musoW dataset, we were capable of identifying a set of unexplored opportunities for Semantic Web technologies. Future work includes the enhancement of the resources descriptions with the results of the analysis, and support the exploration of the dataset with visualizations. For example, we intend to augment the musoW catalogue with a classi cation of the resources with respect to research tasks. Finally, we intend to study how pragmatically the musoW dataset can support musical researchers in the discovery and adoption of Web data, for example linking the collection to prototypical work ows for musical enquiry. 12 W3C PROV-O: https://www.w3.org/TR/prov-o/ 13 Open Annotation Data Model (Community draft): http://www.openannotation.
org/spec/core/