One current challenge in linked science is to adequately describe where a piece of information in the linked science cloud came from. Provenance models, such as Proof Markup Language (PML), have developed methods for expressing simple relationships between information and the sources of information. We argue that the representation of where information comes from is central to trusting linked data in scienti c applications. We introduce the notion of a model of information source and the usage of the source to obtain information by describing the Proof Markup Languages notion of source usage and show how this relationship can be modeled in a library science schema, Functional Requirements for Bibliographic Resources (FRBR). We discuss how these kinds of representations are critical to provenance models.
Before publishing a result, scientists need to check their facts. We stand on the shoulders of giants, but as we push forward in science, we need to make sure that we aren't standing on a giant house of cards. Knowing how, when and where your data comes from is critical for good science, and it's even more critical for linked science, where it isn't immediately clear where a database record or knowledge assertion came from. Sources of information become critical to evaluate information quality. It is di cult if not impossible to assess the trust of information, or to encode it as knowledge, without having a link between information and their sources. For example, one may want to know if the information came from a source such as the New York Times, and further, it may be useful to know the date, edition, page, and exact text fragment where the information was asserted.
There are many challenges in the task of assigning a source to a piece of information. First, it may not be easy to characterize the piece of information in larger information containers (databases, printed documents, web documents, documents that require parameters from a system to be retrieved, etc). Second, the source of a piece of information is often a source of other pieces of information and should be referenced by an identi er and characterized elsewhere. Third, the assertion of the piece of information is a point-time event that occurs during the life-time of the information source. Thus, not any assertion event is a valid assertion: it needs to occur during the lifespan of its source(s) and in places where the sources are located. Those are all critical conditions that need to be properly captured in provenance languages if one wants to make proper use of source information in combination with linked data. 2
The Proof Markup Language (PML) [
An example of this representation can be seen in Figure 11. The top level
concept of a NodeSet supports the encoding of support for a particular piece
of information that can be viewed as a conclusion of some inference step. That
inference step can be as simple as a told assertion or could be an inference using
some antecdents and resulting in a conclusion. One type of inference includes
the usage of a document source to obtain a piece of information. Figure 1 shows
a particular usage of a document (the PML primer) and includes an encoding of
the time it was used and the fragment of the text used. The inference steps use of
a SourceUsage identi es the following: the date, time, location and source of the
assertion. The Source is a pmlp:Source concept that can be used to represent
things like publications, documents, websites, datasets, person, organizations,
etc. Additional examples are available in Murdock et al. [
The PML classes for Source, SourceUsage, and Information are empirically derived, that is, they are responses to a set of use cases that required tracking where information came from and how it was used. This model was e ective 1 The RDF can be downloaded from http://inference-web.org/proofs/csctest/iwppml-2.rdf at capturing the text analytic requirements from the Unstructured Information Management Architecture components, however a more general representation may be bene cial to support additional use cases such as copying les, transforming data from one format to another, and so on. Files on disk are considered to be pmlp:Sources, which limit the ability to describe mechanical duplication, as derivational provenance in PML is limited to pmlp:Information. Similarly, transformation of data from one le format to another results in, in one perspective, two pmlp:Information instances that have the same information (they have the same information), but in one perspective have completely di erent le content, since the information is being represented using a di erent le encoding. Generalization of these sorts of relationships can allow faithful representation of these operations and allow for extension and decomposition of concepts like "the source of a piece of information".
Library Science has spent signi cant time dealing with some kinds of
provenance in the realm of bibliographic resources. Functional Requirements for
Bibliographic References (FRBR) [
Describing how information is derived from a source is a little more complicated in FRBR, but is more generalizable and allows for greater shades of distinction. We propose that, since pmlp:Source is considered to be an opaque, speci c resource, it should be a subclass of frbr:Item. pmlp:Information is actually a role of a frbr:Work, frbr:Expression, or frbr:Manifestation. For instance, an expression may not be information but may play the role of being information in the context of an assertion. In Figure 2 two FRBR stacks show how a quote "This document provides a brief introduction to the Proof Markup Language (PML)" from the PML Primer2 is derived from a downloaded copy of the primer. In the representation, we use the abstractive perspective to allow for description of physical movement of data and transformation of information using the same derivational ontology. Conversely, it allows description of what happened encoded directly in the relationships. For instance, the fact that the copyEvent produced an identical copy is stated by the fact that the server copy and client copy are exemplars of the same Manifestation, while the subset event produces a quote simply because that stack is declared to be partOf the PML Primer stack. The derivational ontology is not named, but PML is an adequate candidate for this task. Its information source construction can be replaced by FRBR. 4
We believe that any serious model of provenance that supports linked science must provide a mechanism for describing information sources and their usage. This can be and has been achieved using the modeling primitives provided in PML. By using the mapping we describe using FRBR, we can also model additional nuanced explanations of data access, transformation, and analysis. Generalized models of abstractive provenance also provide opportunities to express nuanced explanations of data access, transformation, and analysis. We show how the link between information and source can be modeled using a combination of FRBR and a derivational provenance model. This combination is powerful, and allows for unambiguous descriptions of data and information access and transformation. Finally, we argue that the abstractive dimension should be a key component of any provenance model that attempts to deal with artifacts of information or data. 2 http://inference-web.org/2007/primer/
The Tetherless World Constellation is partially funded by DARPA, U.S. Department of Energy, Fujitsu, LGS, Lockheed Martin, Microsoft Research, NASA, National Ecological Observatory Network (NEON), the National Science Foundation, Qualcomm, and the Woods Hole Oceonographic Institution (WHOI). This research was partially funded by the National Science Foundation under CREST Grant No. HRD-0734825.