The OpenBiblio workshop took place on 6th May 2011, at London Knowledge Lab
- Peter Murray-Rust (Open Bibliography project, University of Cambridge, IUCr)
- Mark MacGillivray (Open Bibliography project, University of Edinburgh, OKF, Cottage Labs)
- William Waites (Open Bibliography project, University of Edinburgh, OKF)
- Ben O’Steen (Open Bibliography project, Cottage Labs)
- Alex Dutton (Open Citation project, University of Oxford)
- Owen Stephens (Open Bibliographic Data guide project, Open University)
- Neil Wilson (British Library)
- Richard Jones (Cottage Labs)
- David Flanders (JISC)
- Jim Pitman (Bibserver project, UCB) (remote)
- Adrian Pohl (OKF bibliographic working group) (remote)
During the workshop we covered some key areas where we have seen some success already in the project, and discussed how we could continue further.
Open bibliographic data formats
In order to ensure successful sharing of bibliographic data, we require agreement on a suitable yet simple format via which to disseminate records. Whilst representing linked data is valuable, it also adds complexity; however, simplicity is key for ensuring uptake and for enabling easy front end system development.
Whilst data is available as RDF/XML, JSON is now a very popular format for data transfer, particularly where front end systems are concerned. We considered various JSON linked data formats, and have implemented two for further evaluation. In order to make sure this development work is as widely applicable as possible, we wrote parsers and serialisers for JSON-LD and RDF/JSON as plugins for the popular RDFlib.
The RDF/JSON format is, of course, RDF; therefore, it requires no further change to enable it to handle our data, and our RDF/JSON parser and serialiser are already complete. However, it is not very JSON-like, as data takes the subject(predicate(object)) form rather than the general key:value form. This is where JSON-LD can improve the situation – it provides for listing information in a more key:value-like format, making it easier for front end developers not interested in the RDF relations to utilise. But this leads to additional complexity in the spec and parsing requirements, so we have some further work to complete:
* remove angle brackets from blank nodes
* use type coersion to move types out of main code
* use language coersion to omit languages
Our code is currently available in our repository, and we will request that our parsers and serialisers get added to RDFlib or to RDFextras once they are complete (they are still in development at present).
To further assist in representing bibliographic information in JSON, we also intend to implement BibJSON within JSON-LD; this should provide the necessary lined data functionality where necessary via JSON-LD support, whilst also enabling simpler representation of bibliographic data via key:value pairs where that is all that is required.
By making these options available to our users, we will be able to gauge the most popular representation format.
Regardless of format used, a critical consideration is that of stable references to data. Without this maintaining datasets will be very hard. To date, the British Library data for example does not have suitable identifiers. However, the BL are moving forward with applying identifiers and will be issuing a new version of their dataset soon, which we will take as a new starting point. We have provided a list of records that we have identified as non-unique, and in turn the BL will share the tools they use to manage and convert data where possible, to enable better community collaboration.
Getting more open datasets
We are building on the success of the BL data release by continuing work on our CUL and IUCr data, and also by getting more datasets. The latest is the Medline dataset; there were some initial issues with properly identifying this dataset, so we have a previous blog post and a link to further information, the Medline DTD and specifications of the PubMed data elements to help.
The Medline dataset
We are very excited to have the Medline dataset; we are currently working on cleaning so that we can provide access to all the non-copyrightable material it contains, which should represent a listing of about 98% of all articles published in PubMed.
The Medline dataset comes as a package of approximately 653 XML files, chronologically listing records in terms of the date the record was created. This also means that further updates will be trackable as they will append to the current dataset. We have found that most records contain useful non-copyrightable bibliographic metadata such as author, title, journal, PubMed record ID, and that some contain further metadata such as citations, which we will remove. Once this is done, and we have checked that there are unique IDs (e.g. that the PubMed IDs are unique) we will make the raw CC0 collection available, then attempt to get it into our Bibliographica instance. We will then also be able to generate visualisations on our total dataset, which we hope will be approaching 30 million records by the end of the JISC Open Bibliography project.
Displaying bibliographic records
Whilst Bibliographica allows for display of individual bibliographic records and enables building collections of such records, it does not yet provide a means of neatly displaying lists of bibliographic records. We have partnered with Jim Pitman of Berkeley University to develop his BibServer to fit this requirement, and also to bring further functionality such as search and faceted browse. This also provides further development direction for the output of the project beyond the July end date of the JISC Open Bibliography project.
Searching bibliographic records
Given the collaboration between Bibliographica and BibServer on collection and display of bibliographic records, we are also considering ways to enable search across non-copyrightable bibliographic metadata relating to any published article. We believe this may be achievable by building a collection of DOIs with relevant metadata, and enabling crowdsourcing of updates and comments.
This effort is separate to the main development of the projects, however would make a very good addition both to the functionality of developed software and to the community. This would also tie in with any future functionality that enables author identification and information retrieval, such as ORCID, and allowing us to build on the work done at sites such as BIBKN
Disambiguation without deduplication
There have been a number of experiments recently highlighting the fact that a simple LUCENE search index over datasets tends to give better matches than more complex methods of identifying duplicates. Ben O’Steen and Alex Dutton both provided examples of this, from their work with the Open Citation project.
This is also supported by a recent paper from Jeff Bilder entitled “Disambiguation without Deduplication” (not publicly available). The main point here is that instead of deduplicating objects we can simply do machine disambiguation and make sameAs-ness assertions between multiple objects; this would enable changes to still be applied to different versions of an object by disparate groups (e.g. where each group has a different spelling or identifier, perhaps, for some key part of the record) whilst still maintaining a relationship between the two objects. We could build on this sort of functionality by applying expertise from the library community if necessary, although deduplication/merging should only be contemplated if there is a new dataset being formed which some agent is taking responsibility to curate. If not, better to just cluster the data by SameAs assertions, and keep track of who is making those assertions, to assess their reliability.
We suggest a concept for increasing collaboration on this sort of work – a ReCaptcha of identities. Upon login, perhaps to a Bibliographica or another relevant system, a user could be presented with two questions, one of which we know the answer to, and the other being a request to match identical objects. This, in combination with decent open source software tools enabling bibliographic data management (building on tools such as Google Refine and Needlebase), would allow for simple verifiable disambiguation across large datasets.
Sustaining open bibliographic data
Having had success in getting open bibliographic datasets and prototyping their availability, we must consider how to maintain long term open access. There are three key issues:
Continuing community engagement
We must continue to work with the community, and to provide explanatory information to those needing to make decisions about bibliographic data, such as the OpenBiblio Principles and the Open BIbliographic Data guide. We must also ensure we improve resource discovery by supporting the requirement for generating collections and searching content.
Additionally, quality bibliographic data should be hosted at some key sites – there are a variety of options such as Freebase, CKAN, bibliographica – but we must also ensure that community members can be crowdsourced both for managing records within these central options and also for providing access to smaller distributed nodes, where data can be owned and maintained at the local level whilst being discoverable globally.
Dataset maintenance is critical to ongoing success – stale data is of little use to people and disregard for content maintenance will put off new users. We must co-ordinate with source providers such as the BL by accepting changesets from them and incorporating that into other versions. This is already possible with the Medline data, for example, and will very soon be the case with BL updates too. We should advocate for this method of dataset updates during any future open data negotiations. This will allow us to keep our datasets fresh and relevant, and to properly represent growing datasets.
We must continue to promote open access to non-copyrightable datasets, and ensure that there is a location for open data providers to easily make their raw datasets available – such as CKAN.
We will ensure that all the software we have developed during the course of the project – and in future – will remain open source and publicly available, so that it will be possible for anyone to perform the transforms and services that we can perform.
Community involvement with dataset maintenance
We should support community members that wish to take responsibility for overseeing updating of datasets. This is critical for long term sustainability, but hard to find. These people need to be recruited and provided with simple tools which will empower them to easily maintain and share datasets they care about with a minimal time commitment. Thus we must make sure that our software and tools are not only open source, but usable by non-team members.
We will work on developing tools such as ReCaptcha for disambiguation, and on building game / rank table functionality for those wishing to participate in entity disambiguation (in addition to machine disambiguation).
We hope that by providing almost 30 million records to the community under CC0 license, and with the support of all the providers that made this possible, we will achieve a critical mass of data, and an exemplar for future open access to such data.
This should provide the go-to list of such information, and inspire others to contribute and maintain. However, such community assistance will only continue for as long as there appears to be reasonable maintenance of the corpus and software we have already developed – if this slips into disrepair, community engagement is far less likely.
The bibliographica service that we currently run already requires significant hardware to run. Once we add in Medline data, we will require very large indexes, requiring a great deal of RAM and fast disks. There is therefore a long term maintenance requirement implicit in running any such central service of open bibliographic data on this scale.
We will present a case for ongoing funding requirements and seek sources for financial support both for technical maintenance and for ongoing software maintenance and community engagement.
In order to ensure future engagement with groups and business entities, we must make clear examples of the benefits of open bibliographic data. We have already done some work on visualising the underlying data, which we will develop further for higher impact. We will identify key figures in the data that we can feed into such representations to act as exemplars. Additionally, we will continue to develop mashups using the datasets, to show the serendipitous benefit that increases exposure but is only possible with unambiguously open access to useful data.
Events and announcements
We will continue to promote our work and the efforts of our partners, and advocate further for open bibliography, by publicising our successes so far. We will co-ordinate this with JISC, BL, OKF and other interested groups, to ensure the impact of announcements by all groups are enhanced.
We will present our work at further events throughout the year, such as attendance and sessions at OKCon, OR11 and other conferences, and by arranging further hackdays.