Annotations as Interpretations and Practices

There are usually two ways to look at the consensus represented by a data model or architecture: as the interpretation of a domain and as a plan of action. XML schemas, UML diagrams, SQL relational definitions, RDF vocabularies, and many other familiar tools can be read both as accounts of the entities and relational properties in a domain, and as guidelines for uniformity of naming and encoding. These two understandings are not necessarily at odds with each other: a specification may lay out both how its authors understand what they're modeling and how they have agreed to maintain consistency in expressing their data. But for purposes of promoting adoption, one often finds more relative emphasis on either the social power of uniform practice or on the generality of an elegant declarative account. We may rely on one of these more than the other because it's simply difficult to agree either on how we understand our domains or how we'll standardize our practices (let alone both).

For an example of where we might emphasize one modeling agenda over another, consider the application of encoding machine-readable annotations of digital resources. In an address to a humanities computing symposium at King's College, John Unsworth listed annotation among seven scholarly primitives, i.e., "basic functions common to scholarly activity across disciplines, over time, and independent of theoretical orientation" (Unsworth, 2000). Supporting digital annotation has long been a core use case in the humanities, but the growing role of response and feedback in both online social media and digital commerce has encouraged the understanding of annotations as a resource genre in their own right. A variety of architectures have consequently been adopted or proposed as bases for digital annotation, including HyTime (Calabretto and Pinon, 1997), the Multivalent Framework (Phelps and Wilensky, 1997), the Linguistic Annotation Framework (ISO, 2012) and the Open Annotation Data Model (Sanderson et al., 2013).

Open Annotation (OA) is an RDF vocabulary for relationships among resources, and for metadata about those relationships. OA is developed and published by a confluence of earlier projects now called the Open Annotation Community. The following OA serialization is taken from their cookbook of examples:[1]

Figure 1

ex:Anno a oa:Annotation ;
   oa:hasTarget <http://www.foaffy.org/gallery2/main.php?g2_itemId=11496> ;
   oa:hasBody <http://en.wikipedia.org/wiki/Harry_Perry_%28musician%29> ;
   oa:annotatedBy ex:Person1 ;
   oa:annotatedAt "2012-02-12T15:02:14Z" ;
   oa:serializedBy ex:Software1 ;
   oa:serializedAt "2012-02-12T15:02:14Z"  .

In this example, the body of the annotation is a web page describing the content of an image, the target of the annotation. In OA, an annotation body can be any resource that is somehow "about" another resource, i.e., that stands in some kind of annotating relation to the target. On the one hand, this is a descriptively general account of annotations as a class. But on reflection, OA's generality seems to beg the question of what specific gap it fills in the RDF model of resource associations. If, according to OA, annotations are sets of related resources, what is it that distinguishes an annotation from any other RDF description? OA's answer would seem to be that annotations are exactly those descriptions serialized by the OA community employing the OA vocabulary. But instead of a common vocabulary, why not connect resources in a common architecture across diverse RDF applications? Suppose an annotation vocabulary were linked to specific annotating properties in local languages—something like an ISO 10744 enabling architecture. Properties such as comment_on or reviews could be declared as subproperties of annotation, and resources linked by these annotating relations inferred to be targets and bodies via domain and range declarations as in Figure 2:

Figure 2

  • Equation (a)

    annotatingRelation ⊑ annotates
  • Equation (b)

    ⊤ ⊑ ∀annotates.Target
  • Equation (c)

    ∃annotates.⊤ ⊑ Body

Annotation as a superproperty

This enabling architecture approach offers some advantages. Adopters of such a model would not have to convert their data to a different serialization, for example, and the question of what exactly can be annotated or serve as an annotation body could be governed by range and domain constraints on local subproperties (product reviews annotating products, footnotes annotating essays, etc.). Currently, OACG recommends the Dublin Core Types vocabulary for typing annotation targets and bodies, but the range of the hasTarget and hasBody relations are purposefully undefined. As a result, nothing in the current OA specifications forbids interpretively challenging examples such as a person serving as the annotation body for a target integer or a geographic location as the body for a target person.

On the other hand, the descriptive approach has disadvantages as well. The OA community views annotations as distinct from annotation bodies, but still as individuals with properties of their own. Modeling annotations as relational instances would therefore seem to require reification of either RDF statements or named graphs (or the full existential quantification of first order logic) if annotations are to stand in relations as individuals. So the flexibility of using one's own vocabulary comes at the cost of complexity or a requirement for greater expressive power.

Where do we annotate?

Prescriptive and descriptive emphases can also be contrasted in how particular regions of a resource are targeted for annotation. The OA Specifiers Module recommends particular addressing systems (including XPointer, W3C Media Fragments, etc.) and organizes them by type: fragment selectors, range selectors, and area selectors. But the OA specification does not make clear whether these three categories are exhaustive, or even whether the particular methods listed exhaust the categories (for example, whether there are area selectors in addition to the one listed for SVG files). Figure 3 shows OA's provisional recommendation for combining selectors: they are arranged in an ordered list from most general to most specific. Caveats warn the reader that this chaining method may eventually be deprecated if the RDF community recommends a better approach to expressing order in graph structures[2] .

Figure 3

OA List Composite for chained selectors

Compare the current iterative recommendation in Figure 3 with the (hypothetical) recursive approach shown in Figure 4. This alternative is inspired by HyTime's location ladders, and dispenses with the need for list composites. The targets are organized from specific to general, allowing a more consistent range for the oa:hasSelector property: Figure 3 shows lists in the range, even though lists are not selectors. The point of the example is not that recursion is better than iteration, but that the second solution relies more on a proposed semantics for selectors, while the first appeals to broader community decisions on how best to express ordered relations in RDF.

Figure 4

If OA used HyTime's location ladders

Summary and Generalization: Data Model Roles

Data models and architectures play two kinds of role in information interchange. On the one hand they present a theory of how one should understand the domain of a specification or application, and on the other hand they offer a plan of action. One might call the former role a representational agenda: the interest in doing justice to whatever it is that is being modeled. The second role we'll call a cohortative agenda: promoting uniformity of practice through stipulative definitions.

Successful data models need to address both the representational and cohortative agendas to some extent. Models by their nature are simplifications of reality, and never represent the richness of their domains with full fidelity. Developers adopting a specification must therefore accept its architecture with whatever limitations they find. On the other hand, data models must offer some semantic and expressive adequacy—they won't be adopted if they're not inclusive enough for stakeholders' use cases or if they simply make no sense.

Just as stipulative definitions fill in a model's representational gaps, elegance and descriptive power reduce the need for arbitrary choices. As a result of this trade-off one sees two complementary costs in standards adoption: that of relinquishing full control over stipulations to a broader community vs. the complexity necessary for generality of scope and flexibility of application. But the language of information standards doesn't lay out this complementarity to their audience: RDF schemas and UML diagrams don't announce the degree to which we should read them as micro-theories as opposed to operational definitions.

David Birnbaum presents a vivid illustration of these competing model agendas in his 1996 recommendations for Unicode encoding of early Cyrillic materials (Birnbaum, 1996). Unicode organizes writing systems at the level of the script, rather than the language-specific alphabet or orthography, but specifies that characters convey distinctions in sound or meaning. As Birnbaum points out, orthographic systems that share the same script may assign different sounds or meanings to the same grapheme. The range of languages, regions, and time periods covered in Cyrillic materials scholarship therefore challenges Unicode's representational weakness in ways that typical business uses do not. But, as Birnbaum explains, although distinctions important to Slavic scholars may not always be straightforward to express in Unicode, the "academic and computational chaos and parochialism" that preceded Unicode's introduction lead to even more difficult text rendering and processing problems. Therefore Birnbaum's proposed encoding guidelines for Cyrillic texts address a range of problems for Slavic scholarship, while keeping Unicode's standardization and uniformity. This article is a rare example of making clear both the descriptive limitations of an architecture and the attraction of workarounds for leveraging the power of simply getting on board despite the problems.

Although the Open Annotation specifications don't lay out descriptive/prescriptive tensions as directly, details that may seem at first like semantic infelicities make more sense from a cohortative perspective. OA is primarily a set of encoding guidelines, not a theory of annotation. One adopts the Open Annotation Data Model for the value of participating in a collective enterprise, and reads the specification as a record of the working group's decisions. The promise of OA lies more in prospects for social success than in the model's descriptive power.

Standards development and adoption are social activities—even formal aspects like data modeling. Decisions on which standards to adopt require that we weigh different kinds of costs against each other (e.g., complexity vs. stipulation costs). These decisions take place against a backdrop of social factors that usually aren't documented in standards or discussed in their supporting literature. But each conceptual framework, layered architecture or RDF schema we're offered has something to assert about how things are, and some plan to propose on how to proceed.

Acknowledgments

This research was supported by Libraries: Transformation of Humanities (LG-06-11-0326), an IMLS-funded project focused on building a network of resource collections for the humanities and creating research tools for a wide range of users. The authors thank Karen M. Wickett, Sayeed Choudhury, Gregory Crane, Bridget Almas, Timothy DiLauro, Jim Martino and the anonymous Balisage reviewers for discussions and feedback on earlier versions of this paper.

References

[Birnbaum, 1996] David Birnbaum. Standardizing characters, glyphs, and sgml entities for encoding early Cyrillic writing. Computer Standards and Interfaces, 18:201–252, 1996. doi:https://doi.org/10.1016/0920-5489(95)00096-8.

[Calabretto and Pinon, 1997] Sylvie Calabretto and Jean-Marie Pinon. Modeling a medieval manuscript database with HyTime. In John Smith, editor, Proceedings of ICCC/IFIP Conference on Electronic Publishing: EP'97 New Models and Opportunities, pages 336–345, Canterbury (Great Britain), April 1997. ICCC Press.

[ISO, 1997] ISO. ISO/IEC 10744:1997 Information technology — Hypermedia/Time-based Structuring Language (HyTime). International Organization for Standardization, Geneva, 1997.

[ISO, 2012] ISO. ISO 24612:2012 Language resource management — Linguistic annotation framework (LAF). International Organization for Standardization, Geneva, 2012.

[Phelps and Wilensky, 1997] Thomas A. Phelps and Robert Wilensky. Multivalent annotations. In Research and Advanced Technology for Digital Libraries, pages 287–303. Springer, Berlin Heidelberg, 1997. doi:https://doi.org/10.1007/BFb0026734.

[Sanderson et al., 2013] Robert Sanderson, Paolo Ciccarese, and Herbert Van de Sompel, editors. Open Annotation Data Model. World Wide Web Consortium, 2013.

[Unsworth, 2000] John Unsworth. Scholarly primitives: what methods do humanities researchers have in common, and how might our tools reflect this? Presented at Humanities Computing: formal methods, experimental practice, King's College, London, May 2000.

David Dubin

Research Associate Professor

University of Illinois

David Dubin is a Research Associate Professor at the University of Illinois Graduate School of Library and Information Science in Champaign, IL. David conducts research on foundational issues of information representation and description.

Megan Senseney

University of Illinois

Megan Senseney is Project Coordinator for Research Services at the Graduate School of Library and Information Science, University of Illinois at Urbana-Champaign.

Jacob Jett

University of Illinois

Jacob Jett is a doctoral student and Visiting Project Coordinator at the University of Illinois Graduate School of Library and Information Science.