Wikipedia:Wikipedia Signpost/2012-07-02/Op-ed


 * Neil Jefferies is the research and development project manager at the Bodleian Libraries. He was involved with the initial setup of the Eprints and Fedora Repositories at the University of Oxford, and is now working on the implementation of a long-term digital archive platform. He is the technical director of Cultures of Knowledge, a collaborative project launched in 2009 "to reconstruct the correspondence networks central to the revolutionary intellectual developments of the early modern period".

This piece examines a key question that new Wikimedia projects such as Wikidata are concerned with: how to properly represent knowledge digitally at the most basic level. There is a real danger that an inflexible, proscriptive approach to data will severely limit the scope, capabilities and ultimate utility of the resulting service.

At one level, the textual representation of information and knowledge in books and online can be viewed as simply another serialisation and packaging format for information and knowledge, optimised for human rather than machine consumption. Within the Wikipedia community – Wikidata and elsewhere – there is a perceived utility in using more structured, machine-friendly formats to enable better information sharing and computer-assisted analysis and research. However, there remains a lot of debate about the best approach, to which I will contribute the views I have developed over nearly a decade of research and development projects at the Bodleian Library and before that, through my involvement with knowledge management in the commercial domain.

My first point is that metadata and data are really different aspects of a continuum. In the majority of cases, data acquires much of its meaning only in connection with its context, which is largely contained within so-called metadata. This is especially true for numerical data streams, but holds even for data in the form of text and images: when and where a text was written are often critical elements in understanding the meaning. Data and metadata should be considered not as distinct entities but as complementary facets of a greater whole.

Secondly, there will be no single unifying metadata "standard" (or even a few such standards), so deal with it! For example, biosharing.org lists just under 200 metadata standards for experimental biosciences alone. The notion of a single standard that led to the development of MARC, and latterly RDA, in the library sphere is simply not applicable to the way in which metadata is now used within the field of academic enquiry. This means that any solution to handling digital objects must have a mechanism for handling a multiplicity of standards, and ideally within an individual object – for example, bibliographic, rights and preservation metadata may quite reasonably be encoded using different standards. The corollary of this is that if we have such a mechanism there is no need to abandon existing standards prematurely. This avoidance of over-proscribing and premature decision-making will be familiar to Agile developers. Consequently, Wikidata developers would be ill-advised to aim for a rigid, unitary metadata model – even at a basic level, representing knowledge is too complex and variable for such an approach.

So how do we balance this proliferation of standards with the desire for sharing and interoperability? We can find several key areas in which a consensus view is emerging, not through explicit standard-setting activities but through experience and necessity. This gives us a good indication that these are sensible points on which to base longer-term interoperability.


 * 1) An emergent data/object model. Besides the bibliographic entities, such as digitised texts, images and data, a number of key types of "context-object" recur when we start to try to build more complex systems for handling digital information. This can be seen in such diverse areas as the specifications for TEI, Freebase, CERIF and schema.org. The most important of these elements are people, places, vocabularies/ontologies and the notion of time dependency. Indeed, for many projects in the humanities, these objects actually form the basis for expressing ideas and framing discourse using the conventional bibliographic objects to provide an evidentiary base.
 * 2) Aggregations as a key organising tool for this expanded universe of digital objects. In many cases, these aggregations are also objects in their own right, representing content collections, organisations, geopolitical entities and even projects – each potentially with a history and other attributes. An essential characteristic of aggregations is that they need not be hierarchical, but rather a graph capable of capturing the more unstructured, web-like way people have of organising themselves and their knowledge.
 * 3) Agreement on essential common properties. For each object type there is usually a general consensus on a minimal set of properties that are sufficient to both uniquely identify an object and provide enough information to a human reader that the object is the one that they are interested in. Often, the latter is actually a less strict requirement as a person can use circumstantial evidence such as the context in which an object occurs for disambiguation. While it is desirable to try to capture contextual information systematically, we have to accept that this is frequently not done. Sources for this common baseline could include Dublin Core (or dcterms to be explicit) records, DataCite records, gazetteers, and name authority lists, for example.

These common properties are obviously very amenable to storage and manipulation in a relational database. Indeed, for large-scale data ingestion with the following clean-up, de-duplication and merging of records/objects, this is likely to be the best tool for the job. However, once this task has been completed and we delve into the more varied elements of the objects, the advantages of a purely relational database approach are less clear-cut.

Instead, we can treat each object as an independent, web-addressable entity – which in practice is desirable in its own right as a mode of publication and dissemination. In particular, we can use search engines to index across heterogeneous fields – Apache Solr excels at faceting and grouping, while ElasticSearch can index arbitrary XML without schemas (i.e. all of the varied domain-specific metadata). These tools give users ways into the material that are much easier to use and more intuitive.

The objects alone are only a part of the picture – the relationships between objects are critical to the structure of the overall collection. In fact, in many cases (especially in the humanities) a significant proportion of research activity actually involves discovering, analysing and documenting such relationships. The Semantic Web or, more precisely, the ideas behind the Resource Description Framework (RDF) and linked data, provide a mechanism for expressing these relationships in a way that is structured, through the use of defined vocabularies, but also flexible and extensible, through the ability to use multiple vocabularies. While theoretically it is possible to express all metadata in RDF, this is not practical for performance and usability reasons, and is unnecessary.

This model of linked data, combining a mix of standardised fields and less-structured textual content, should not be entirely unfamiliar to people used to working with Semantic MediaWiki, sharing their metadata on Wikidata, or using data boxes in Wikipedia! However, when applying this model to practical research projects it emerges that a critical element is still lacking. Although we can describe relationships between objects using RDF, we are limited to making assertions of the form  (the RDF "triple"). In practice, relatively few statements of this form can be considered universally and absolutely true. For example: a person may live at a particular address but only for a certain period of time; the copyright on a book may last for 50 years, but only in a particular country. Essentially, what is needed is a mechanism to define the circumstances under which a relationship can be considered valid. A number of possible mechanisms could do this – replacing RDF triples with "quads" that include a context object; annotation of relationships using OAC.

These examples are really just special cases of a more general requirement that is of great interest to scholars. This is the ability to qualify a relationship or assertion to capture an element of provenance. Specifically, we need to know who made an assertion, when, on the basis of what evidence, and under which circumstances it holds. This may be manifested in several ways:


 * Differences of scholarly opinion – it should be possible for there to be contradictory assertions in the data relating to an object, provided we can supply the evidence for each point of view.
 * Quality of the evidence – information can be incomplete, or just unclear if we are dealing with digitised materials. In this case we want to capture the assumptions under which an assertion is made.
 * Proximity of evidence – we may have an undated document but if we know the biography of the author we can place some limits on probable dates. This evidence is not intrinsic to the object but can be derived from its context.
 * Omissions – collections are usually incomplete for various reasons. It is important to distinguish the absence of material as a result of inactivity or specific omission from subsequent failures in collection building.

These qualifications become especially important when we try to use computational tools such as analytics and visualisation. Indeed, projects such as Mapping the Republic of Letters (Stanford University) are expending significant effort to find ways of representing uncertainty and omission in visualisations.

I believe there needs to be a subtle change in the mindset when creating reference resources for scholarly purposes (and, arguably, more generally). Rather than always aiming for objective statements of truth we need to realise that a large amount of knowledge is derived via inference from a limited and imperfect evidence base, especially in the humanities. Thus we should aim to accurately represent the state of knowledge about a topic, including omissions, uncertainty and differences of opinion.