How the IASSIST 2017 keynote framed data as a shared language

In May 2017, librarians, data stewards, and domain scientists gathered at the University of Kansas in Lawrence for IASSIST 2017, a conference built around the theme "Data in the Middle: The Common Language of Research." The four-day programme ran from 23 to 26 May and brought together practitioners working with social, environmental, and agricultural data. Plenary sessions sat alongside hands-on workshops, and the keynote became a touchstone for conversations that continued well after delegates returned home.

The central argument of that keynote was straightforward but quietly radical: data, when properly curated, described, and shared, functions as a common language across disciplines that otherwise struggle to talk to one another. Sociologists, agronomists, climate modellers, and historians rarely share methods, but they can share datasets. That simple insight reframed how delegates thought about metadata, repositories, and the slow, unglamorous work of data management.

For audiences watching from the southern hemisphere, and particularly for those in research-intensive institutions across Australia, the keynote arrived at a moment of growing investment in national data infrastructure. Australian researchers were already navigating the country's complex federated data landscape, balancing Commonwealth privacy laws with state-based health and education datasets. The Lawrence conversation offered a vocabulary that felt immediately useful, and the conference's collegial tone sat well with the Australian preference for consensus-driven exchange, the kind that happens as readily over a flat white in a Melbourne laneway as in a formal seminar room.

What follows is an exploration of the keynote's core ideas, the way deep learning entered the programme, and the Australian echoes of the themes raised in Kansas. It also looks at the practical steps Australian institutions might take if they want to treat data as a shared language rather than a series of isolated dialects.

The Lawrence gathering and the case for a common vocabulary

IASSIST has long been the professional home of data librarians and social science data specialists, but the 2017 programme deliberately widened the tent. Sessions covered digital agriculture, deep learning, and global food security alongside the more familiar territory of survey archives and citation management. The conference organisers wanted to demonstrate that the techniques developed for handling census microdata could be transferred to sensor streams, satellite imagery, and genomic repositories.

The keynote speaker framed data as a working language rather than a static artefact. Just as spoken languages carry grammar, idiom, and context, datasets carry schemas, units, provenance, and licensing. Two researchers can read the same CSV file and understand entirely different things if their vocabularies of variable names and code values differ. The talk argued that the work of building shared metadata standards is, in effect, the work of building a shared grammar.

This framing resonated strongly with delegates from agricultural research backgrounds, who have long wrestled with the compatibility problem in soil science, crop trials, and livestock monitoring. A wheat trial in New South Wales and a barley trial in Western Australia may both record "yield," but the underlying units, plot sizes, and seasonal adjustments can vary enough to make naive comparison misleading. The keynote's call for shared vocabularies, persistent identifiers, and machine-readable provenance felt like an answer to a problem many attendees had been living with for years.

How researchers across disciplines began speaking data

The shift toward data as a shared language did not happen overnight. It grew out of the open science movement, the rise of repositories such as Dryad and the Australian Data Archive, and the gradual adoption of data management plans by major funders. By 2017, the conversation had moved from whether to share data to how to share it in ways that preserved meaning.

Disciplines that once guarded their datasets jealously began to recognise that locked data is often unused data. The Australian Bureau of Statistics has spent years building the DataLab environment, which lets approved researchers query sensitive microdata without downloading it. Similar models are appearing in the United Kingdom and Canada, and the IASSIST keynote highlighted these as examples of what disciplined data sharing can look like when privacy, transparency, and reproducibility are all taken seriously.

Universities have responded by creating dedicated data steward roles, sometimes embedded in libraries, sometimes in research offices. At the University of Melbourne and the University of Queensland, data stewards now sit alongside liaison librarians, helping grant applicants write compliant plans and guiding them toward appropriate repositories. The growth of these roles is one of the clearest signs that data has indeed become a working language requiring interpreters and tutors.

Deep learning and the promise of precision agriculture

One of the most discussed plenary sessions at Lawrence focused on the intersection of deep learning and agriculture. The presentation, captured in the conference archive under the headline deep learning for precision agriculture, described how convolutional neural networks trained on drone and satellite imagery could detect crop stress weeks before it became visible to the human eye. The work drew on field trials from the United States Midwest, but the methods translate directly to the wheat belts of southern Australia and the mixed farming systems of the Murray-Darling region.

Precision approaches are already familiar territory for Australian agriculture. GrainCorp, the major bulk handler operating across New South Wales, Victoria, and Queensland, has invested heavily in sensor networks at receival sites, and startups in Adelaide and Perth are building machine learning tools to optimise irrigation in horticulture. What the IASSIST session added to this picture was a vocabulary for talking about model provenance, training data bias, and reproducibility, concerns that have often been afterthoughts in agricultural machine learning.

A few contrasts between traditional agronomic research and the data-driven approaches showcased at the conference are sketched below. They are not a comprehensive map, but they capture the kinds of shifts delegates were being asked to consider.

Aspect	Traditional agronomic research	Data-driven precision agriculture
Primary data source	Field notebooks, lab assays	Sensors, drones, satellite imagery
Time to insight	Seasons to years	Hours to weeks
Unit of analysis	The plot or paddock	The pixel or the animal
Skills required	Botany, soil science, statistics	Statistics, software engineering, domain knowledge
Sharing norms	Journal articles, informal networks	Repositories, APIs, model cards
Reproducibility challenge	Lost notebooks, undocumented methods	Opaque models, biased training data

The comparison matters because it shows where the work of building a shared language actually happens. It is not in the flashy model architecture but in the unglamorous decisions about variable names, units, and documentation. A neural network trained on inconsistently named columns will reproduce inconsistency at scale, and the cost of cleaning data downstream is far higher than the cost of agreeing on a schema upstream.

Open infrastructure, standards, and international collaboration

No language survives without shared standards, and the IASSIST 2017 programme gave significant airtime to the bodies that maintain those standards. Sessions covered schema.org for dataset markup, DataCite for persistent identifiers, and the Research Data Alliance's working groups on metadata. The conference also highlighted the role of instrumentation consortia in making sure that the raw readings flowing into models are themselves comparable across laboratories and borders.

International coordination matters because research problems do not respect national boundaries. A drought in the Mallee region of Victoria has analogues in southern Spain, the American Great Plains, and the African Sahel. Datasets from each region only become useful for comparative work if the variables describing soil moisture, crop variety, and weather have been recorded in compatible ways. Bodies such as the international standards consortium that steward instrumentation protocols were cited as quiet enablers of the kind of cross-border analysis that policy makers increasingly demand.

Australia has its own contributions to make in this space. The Terrestrial Ecosystem Research Network operates long-term monitoring sites from Cape Tribulation to Tasmania, and its data portal is recognised internationally for rigorous metadata. CSIRO's data access governance frameworks have also been studied by overseas agencies trying to balance openness with the privacy expectations of Australian taxpayers, who fund much of the research through the Commonwealth Scientific and Industrial Research Organisation.

What Australian institutions heard in the keynote

For Australian delegates, the Lawrence keynote landed in the middle of a domestic conversation about research data infrastructure. The Australian Research Data Commons had recently launched several major programmes, and the National Data Commissioner was beginning to outline how the Data Availability and Transparency Act would shape access to public sector data. Both initiatives depend on a workforce fluent in the language of data, and both have struggled with the sheer breadth of skills required.

Universities in Sydney, Brisbane, and Perth were already experimenting with different models. Some have built central data teams within the library; others have distributed data stewards across faculties. The keynote suggested that the choice of model matters less than the underlying commitment to treating data as a scholarly output deserving the same care as a published paper. Several Australian universities have since adopted data citation policies that mirror their existing thesis deposit rules.

There are also cultural and legislative specifics that shape how Australian institutions hear the shared language argument. The Privacy Act 1988 imposes obligations that differ from those in the United States, and the Australian Code for the Responsible Conduct of Research sets expectations about data management that researchers must meet to remain in good standing. Building a common vocabulary that respects these local requirements, while still allowing international collaboration, is one of the practical challenges the keynote implicitly raised.

Practical steps for Australian research teams

Translating the IASSIST 2017 themes into local practice does not require new technology so much as new habits. The Australian research sector already has the infrastructure; what it needs is consistent adoption and cultural reinforcement. The following recommendations are drawn from the conference archive and from subsequent Australian initiatives that have tried to keep the shared language metaphor alive.

• Audit existing datasets for variable naming consistency across research groups, and document the choices in a shared style guide.
• Adopt persistent identifiers for datasets, people, and organisations, and cite datasets in publications the way journals are cited.
• Build relationships with domain-specific international standards bodies to ensure that locally collected data can join global comparators without expensive retrofitting.
• Train early career researchers in data management planning as a core research skill, on par with literature searching and statistical analysis.
• Engage with the Australian Research Data Commons and the National Data Commissioner's office to align institutional practice with national legislation and infrastructure investment.
• Treat documentation as part of the research output, and allocate time and recognition for it in promotion and grant assessment criteria.