Assessing the quality of linked bibliographic data

2019-04-02

First, I want to acknowledge that McGill University is located on land which has long served as a site of meeting and exchange amongst Indigenous peoples, including the Haudenosaunee and Anishinabeg nations.

To situate myself, my family is of English and Irish settler origin, and my mother's side traces our arrival to the New York region of North America in the 1700's, before crossing into southern Ontario as part of the Loyalist movement. Since 2005 I've lived in Sudbury, Ontario, which is the territory of the Atikemkshing Aniishnabek in the Robinson-Huron Treaty territory, where I have worked as a systems librarian at Laurentian University.

That last piece partially explains why I'm interested in data quality of linked bibliographic data.

Bibliographic data	Linked data
Printed cards	A web of documents
Digital card catalogues	A web of data
Records (fields with values)	Entities and relationships
Text	Text and media
Records are shared, modified, reshared	Anyone can say anything about anything
Metadata	Metadata and data

Part of a MARC21 record showing fields and text — MARC record data

Corresponding data as BIBFRAME showing relationships between the book and film — Linked data

Linked bibliographic data

From bibliographic records to linked bibliographic data:
- Dublin Core Metadata Terms (dcterms) - 2003-
- Bibliographic Ontology (bibo) - 2008-2009
- FRBR-Aligned Bibliographic Ontology (fabio) - 2010-
- Bibliographic Framework (BIBFRAME) - 2012-
- Resource Description Access (RDA) - 2014-

dcterms is used extensively, both as the basis of other vocabularies, and as a lightweight vocabulary for institutional repositories and other special collections.

While bibo and fabio were developed relatively early, and have influenced other vocabularies such as schema.org, they have found low adoption rates.

BIBFRAME is an initiative led by the Library of Congress to provide a successor for the 50-year old MARC format for bibliographic interchange. The second iteration of BIBFRAME in 2016 saw an international set of libraries including Harvard, Stanford, Cornell, and the University of Alberta commit to adopting BIBFRAME as their primary cataloguing format.

RDA is an effort to produce a vocabulary aligned with FRBR (now IFLA LRM) that targets all cultural heritage institutions.

Data quality: why?

Standards are not sufficient, nor standard (Zhu et al, 2016)
Bibliographic data: history of collaborative efforts and challenges
Linked data: broad scope, sophisticated models, but limited results
Linked bibliographic data: early days, wide open

Seminal definition of data quality

data that are fit for use by data consumers Wang & Strong (1996)

"data consumers" means context is critical
Context for this work was MIS with defence systems

Wang & Strong (1996) model

Category	Dimensions
Intrinsic	believability, accuracy, objectivity, reputation
Contextual	value-added, relevancy, timeliness, completeness, amount
Representational	interpretability, ease of understanding, representational consistency, concise representation
Accessibility	accessibility, access security

Bibliographic data models

Paper	Categories (# dimensions)
Moen et al. (1998)	Completeness, accuracy, serviceability
Bruce & Hillman (2004)	Completeness, accuracy, provenance, conformance, coherence, timeliness, accessibility
Stvilia et al. (2007)	Intrinsic (9), relational/contextual (12), reputational (1)
Ochoa & Duval (2009)	Bruce & Hillman (2004)
Gavrilis et al. (2015)	Completeness, accuracy, consistency, appropriateness, auditability

Linked data models

Paper	Categories (# dimensions)
Pipino et al. (2002)	Wang & Strong (1996) + free-of-error, objectivity
Fürber & Hepp (2011)	Completeness, accuracy, timeliness, uniqueness
Zaveri et al. (2016)	Intrinsic (5), contextual (4), representational (4), accessibility (5)
Färber et al. (2017)	Intrinsic (3), contextual (3), representational (2), accessibility (3)
Debattista et al. (2018)	Intrinsic, contextual, representational, accessibility

Gaps and opportunities

Generalizability

Linked data: yes, but specificity is lacking
Bibliographic literature: no
- Small scale
- Dublin Core
- Context bound

Reproducibility

External validity threat: history vs. open datasets
Lack of access to implementations, in-house datasets

Automation: breadth vs. depth

Billions of records: mostly syntax
Handfuls of records: rich semantics

Research questions

How should libraries conceive of quality in the domain of linked bibliographic data?
How can libraries meaningfully assess the quality of linked bibliographic data?
How can libraries assess the quality of linked bibliographic data at scale?

References

Appendices

Proposal

Synthesize a conceptual model for quality assessment of linked bibliographic data
Compile a gold standard dataset of linked bibliographic data using the BIBFRAME vocabulary (Aalberg et al., 2018; Decourselle, 2016)
Implement the conceptual model metrics in an automated assessment tool

Significance

Constructivist grounded theory is a new approach to model development for data quality
Gold standard dataset will ease the efforts of other researchers, in data quality and other inquiries
Assessment tool should enable libraries to evaluate and improve their linked bibliographic data early

Methodologies

Surveys
Interviews
Card sorting
Content analysis
Log analysis
Functional requirements heuristics
Deductive reasoning
Automated tools

Aggregations

From national union catalogues, cooperative cataloguing efforts, and OAI-PMH harvested records, to open knowledge graphs

Boissonas (1979); Bruce & Hillman (2004); Debattista et al. (2018); El-Sherbini (2010); Färber et al. (2017); Hider & Tan (2008); Moen et al. (1998); Stvilia et al. (2007);

Fitness for use

Heavy dependence on IFLA LRM (2017) tasks:
- Find
- Identify
- Select
- Obtain
- Explore

Data consumers

"Data consumers" depend on context
Linked data broadens possible contexts
Consumers... but also producers?