Conference PaperPDF Available

Tracking the Functions of AI as Paradata & Pursuing Archival Accountability

Authors:

Abstract

While a familiar term in fields like social science research and digital cultural heritage, ‘paradata’ has not yet been introduced conceptually into the archival realm. In response to an increasing number of experiments with machine learning and artificial intelligence, the InterPARES Trust AI research group proposes the definition of paradata as ‘information about the procedure(s) and tools used to create and process information resources, along with information about the persons carrying out those procedures.’ The utilization of this concept in archives can help to ensure that AI-driven systems are designed from the outset to honor the archival ethic, and to aid in the evaluation of off-the-shelf automation solutions. An evaluation of current AI experiments in archives highlights opportunities for paradata-conscious practice
Tracking the Functions of AI as Paradata & Pursuing Archival
Accountability
Jeremy Davet; School of Information, University of British Columbia; Vancouver, Canada
Babak Hamidzadeh; College of Information Studies, University of Maryland; College Park, MD
Patricia Franks; San Jose State University; San Jose, CA
Jenny Bunn; The National Archives; London, United Kingdom
Abstract
While a familiar term in fields like social science research and
digital cultural heritage, ‘paradata’ has not yet been introduced
conceptually into the archival realm. In response to an increasing
number of experiments with machine learning and artificial intel-
ligence, the InterPARES Trust AI research group proposes the
definition of paradata as ‘information about the procedure(s) and
tools used to create and process information resources, along with
information about the persons carrying out those procedures.’ The
utilization of this concept in archives can help to ensure that AI-
driven systems are designed from the outset to honor the archival
ethic, and to aid in the evaluation of off-the-shelf automation solu-
tions. An evaluation of current AI experiments in archives highlights
opportunities for paradata-conscious practice.
Introduction
As machine learning algorithms continue to grow more accu-
rate, adaptable, and affordable, the moment at which artificial intell-
igence is formally implemented in archives draws correspondingly
near. In recent years archives of all stripes have begun experiment-
ing with different AI-powered software, often in the hope of auto-
mating the more menial and time-consuming elements of appraisal,
selection, description, and arrangement. In reviewing several such
experiments, it has become clear that the functioning of AI is as yet
too opaque to meet the bar of accountability and transparency set for
archives by the publics they serve.
The primary goal of this piece is therefore to explore the value
and feasibility of documenting and preserving the procedural func-
tioning of AI as it is applied to data and records. In order to properly
conceptualize this sort of documentation, it is necessary to introduce
the term ‘paradata’ into the archival vocabulary, drawing from the
fields of cultural heritage, social science research, and archaeology.
Defined within the InterPARES Trust AI research group as infor-
mation about the procedure(s) and tools used to create and process
information resources, along with information about the persons
carrying out those procedures, this piece will highlight some oppo-
rtunities to collect and organize ‘paradata’ in the context of the
application of AI to archives.
In this context of collection, paradata can be used to allow aud-
itors, archivists, and/or members of the public to identify functional
weaknesses in automated systems. In so doing, it should facilitate
experimentation with those systems so that they might be made
more efficient or responsive to the specific needs of archives users.
More broadly, it may also be used to ensure that the archives remains
an accountable institution even when certain archival functions are
mediated by automated systems, which themselves cannot be sub-
ject to the same standards of scrutiny as human beings.
While these objectives are all aspirational and supposed upon
the current, developing InterPARES definition, they nonetheless
reflect the spirit in which the concept was developed. The ends to
which paradata are applied are ultimately the purview of the arch-
ives which have collected them. However, the general character of
paradata is use- and user-agnostic.
As current experiments with automation illustrate, there is a
growing need to collect and disseminate information about AI aug-
mented processes in language suitable for archivists, administrators,
programmers, and members of the public alike. Furthermore, those
efforts which address this need must remain mindful of changing
emphases in archival ethics, as well as the current capabilities of AI
and their designers to articulate their workings.
Conceptual Origins and Development
While the intellectual origins of the concept of paradata as
applied to the aforementioned fields of cultural heritage, social sci-
ence research, and archaeology are diffuse and date at least as far
back as 1989, the term itself is generally attributed to a presentation
given by the sociologist Mick P. Couper at the 1998 Joint Statistical
Meeting in Dallas, Texas.[1] Originally used to refer to data created
as a byproduct of automated systems used during the research pro-
cess,[2] paradata has since been generalized outside of the archival
context to also mean information about human processes of under-
standing and interpretation, unintentionally created but nonetheless
instructive and analytically useful in its own right.[3], [4] The lack
of a singular definition of paradata reveals its multifocal application.
Depending on the field, paradata has been used to “[communicate]
uncertainties and the different phases of the process of interpretation
that were often impossible to discern [otherwise]”;[5] to improve
social science survey design, especially to improve response rate
and quality; and to capture information about the researcher-subject
relationship, aiding analysis of the conclusion-forming process.[6]
In every context where it has been applied thus far, paradata
has been used to pursue intellectual accountability, assure oper-
ational transparency, and enable review of important intellectual
decisions; the introduction of the concept of paradata into the arch-
ival sphere is intended to support the pursuit of similar goals.
Relation to Archival Theory
The development of the archival definition of paradata is a
reflection not only of a particular preoccupation with the functioning
of automated systems within archives, but also as an extension of
evolving ideas around archival transparency, accountabilities, and
https://doi.org/10.2352/issn.2168-3204.2022.19.1.17
This work is licensed under the Creative Commons Attribution
4.0 International License.
Contributions to this paper made by Jenny Bunn of
The National Archives of the UK are © Crown copyright
and are re-used under the terms of the Open Government Licence.
ARCHIVING 2021 FINAL PROGRAM AND PROCEEDINGS 83
record integrity. Archivists and archives administrators have in-
creasingly been interested in enumerating their positionalities and
redressing their personal biases, in service of improving the function
of the archives.[7] Some archives have also taken to including
contextual metadata in their descriptive schema, explaining what
interpretations an archivist has made of a record and serving to und-
erscore that these interpretations are mediated, subject to revisitation
and revision.
Paradata seeks to serve a similar but distinct purpose; to define
the steps in and character of the process by which these interpre-
tations were synthesized. The same motivations which have infor-
med the use of positionality statements and contextual metadata
have shaped the conceptualization of paradata; namely, refinements
to archival ethics and an expansion of the publics which archives are
to serve. Moreover, it is intended to fill a perceived void in the crit-
ical evaluation of archival arrangement and description, especially
in cases where the archival agent cannot be conversationally inter-
rogated. By collecting, preserving, and evaluating the information
that said agent used to inform their final interpretation, the func-
tional transparency of the archives is improved, and new avenues by
which archival decisions may be redressed are opened.
While this work and the current research of the InterPARES
Trust AI group emphasizes the relationship of paradata and AI-
enabled automated systems, it should also be recognized that this is
not a limitation inherent to the concept. Rather, it can be applied in
any case where a person or intelligence makes archival decisions in
the normal course of archives’ operations or during periods of exper-
imentation.
Application to Artificial Intelligence (AI)
When applied to the AI-powered systems with which archives
are currently experimenting, the concept of paradata is intended to
describe the often-murky internal operations of these automated sol-
utions. Regularly these programs or systems are in part or in whole
‘black box,’ meaning that the methods by which they interpret an
input and return an output are obscured from the user.[8] The rea-
sons for the development of these black-box solutions are numerous
and range from the particular competencies of different program-
mers to the financial resources of the contracting party. The means
by which they come about are also multitudinous and include every-
thing from the cobbling together of many different programs of
diverse origins to complete complex tasks, to the software automati-
cally self-iterating as more data is input. Whatever the reasons for
the inclusion of black-box elements in a particular AI-utilizing sys-
tem, their ubiquity is undeniable.
This reality has already begun to inform experimentation on the
part of public institutions like DARPA and private concerns
pursuing projects in ‘explainable AI’ (XAI) and leading directly to
the development of tools such as model cards and impact assess-
ments.[9] Conscious of this experimentation, paradata can and
should be used to frame archivists’ understanding of how the pro-
cesses of AI can be explained, and of what information needs to be
collected in order to do so. None of the aforementioned projects
have been specifically created with archivists in mind, and as such
it is incumbent upon archivists to use a concept like paradata to ar-
ticulate their needs to AI developers vis-à-vis the archival mission
and ethic.
Of course, archives and archivists are not always in a position
to dictate the form of the technologies upon which they rely. In the
(likely many) cases where archivists will have to use an off-the-shelf
AI solution, paradata may also be employed to evaluate deficiencies
in the function and output of these tools, identifying areas in which
automated work must be hand-corrected and where systems may be
improved. As an example, the examination of linked training ob-
jects within a corpus of paradata might reveal that the corresponding
AI system was not exposed to a particular file type during training,
explaining why it fails to categorize that file type correctly in prac-
tice.
Ambiguities & Directions for Development
The primary ambiguity embodied in the current definition of
paradata is that of the extent of information collected and classified
under that heading. The exact amount of information necessary to
fully understand the process of archival interpretation will likely
differ case-to-case, and so it is difficult to articulate the full extent
of what might be included in a corpus of paradata absent specific
knowledge of an archive’s operations.
With respect to AI, the rapid development of new AI capabil-
ities and the continual creation of composite AI-powered systems of
archival automation make it difficult to generalize about what infor-
mation needs to be collected to understand their ‘interpretations.’
As will be articulated in the following section, current archival
projects using AI and proposed AI-recording technologies broadly
suggest the sorts of information that it is currently possible to coll-
ect. However, whether this information is suitable to fully explicate
the interpretive process or whether it is feasible to collect this infor-
mation in all cases is still a matter of debate.
It may be that the archival community eventually decides upon
and supports a unified AI-based automation system or agrees upon
shared structural and reporting standards for their individual sys-
tems. The more general conception of paradata should be used to
support these sorts of decisions, and in the event that such unified
systems or shared standards are developed it may be possible to fur-
ther iterate upon the concept of paradata in relation to AI.
Paradata Collection
To a large extent, what might be collected as paradata is already
generated incidentally or collected as a matter-of-course during the
testing and deployment of AI-based automated systems. Data sets
used to train AI models represent an exceedingly common form of
what can be collected as paradata, although the framing of them as
such is novel. As they directly influence the capabilities and comp-
etencies of AI, they can be used to explicate the interpretations of
that AI as paradata.
However, not all AI-enabled systems may have accompanying
training sets; in some cases, the training data may be inaccessible,
lost, or be restricted by legal contracts or security measures. Indeed,
what can be collected, grouped, interpreted, and disseminated as
paradata in any given archival context is reliant on an expansive set
of factors which are impossible to explore fully over the length of
this work.
In order to begin to understand what might be collected as para-
data, some of the practical factors which influence what it is possible
to collect and use, and current efforts to explicate the function of AI
in archives, consider two relatively recent experiments.
84 SOCIETY FOR IMAGING SCIENCE AND TECHNOLOGY
In Codice Ratio
Based at the Roma Tre University and in collaboration with the
Italian National Research Council (CNR), Vatican Apostolic Arch-
ive, and the State Archive of Rome, In Codice Ratio (ICR) was an
experiment launched in 2016 to attempt to improve OCR processing
of medieval manuscripts held by the Vatican.[10] Beginning with
pages scanned from manuscripts in the Vatican Apostolic Archives,
team members from ICR initially relied on dozens of high-school
age volunteers to identify and group penstrokes as letters during the
development of the training data. Eventually, computer scientists
working on the project were able to substitute real images of letters
with procedurally-generated facsimiles for the purposes of training
their AI, and develop systems to segment handwritten words auto-
matically during the preprocessing phase. Once these words were
separated into their constituent letters, they were processed through
a deep convolutional neural network (CNN) that identified letter
type, the extent of a word, and the likeliest translation based on stat-
istical linguistic models.[11] Publishing the bulk of their findings
in 2019, the ICR automated system was eventually able to achieve
a 96% success rate in parsing the contents of medieval manuscripts.
The final documentation of the ICR project includes many
elements which could be classified as paradata and arrives at con-
clusions which were unknowingly informed by said paradata. Most
obvious is their training data set, which they have published in its
entirety on their project website. In evaluating the training set as
paradata, one can understand the bounds of the training strategy
undergirding the ICR project, as well as the functional limitations
that those bounds inspired in the mature ICR system. Specifically,
ICR relied on a training set using only one style of medieval hand-
writinga derivation of the Caroline style.[12] Were the system to
be applied to manuscripts from the same period wrought in a
different writing style, it is likely that its successes would have been
much fewer in number. Additionally, the training data includes
1,000 ‘non-character’ marks, presumably used to train the CNN to
avoid the erroneous classification of letter pairs and flourishes as
discrete letters. One familiar with medieval manuscripts might use
this information in combination with a knowledge of scribes’ fond-
ness for annotation and abbreviation marks to explain the bulk of the
failures that the ICR solution manifested. Due to the sheer variety
of these sorts of non-character marks, a training set of 1,000 was
inadequate to train the system to handle all such cases accurately.
In addition to the training data, other elements might have been pres-
erved from the system as paradata, (Fig. 1) including but not limited
to the source code of the different iterations of the CNN and design
documentation.
The explanations for and conclusions about the function of the
CNN as presented in the numerous journal articles published by the
ICR team were also informed by what can be classified as paradata.
Explanations of how the system segmented each word into letters
would have been informed by design documentation and diagnostic
information generated by said system. Conclusions about the class-
ification strengths and weaknesses inherent to the system would
have relied on performance data and analysis of the training sets.
These can all be considered paradata under the Inter-PARES def-
inition, insofar as they speak to the automated procedures and AI
tools used to create and process information resources.
The National Archives of the UK
Spurred by 2010 changes to the Public Records Act of 1958
newly mandating that the transfer of public records for permanent
preservation should happen no later than 20 years from the date of
creation (formerly 30 years), the National Archives of the United
Kingdom (TNA) redoubled its efforts to evaluate the treatment of
electronic records and to conduct experiments into their automatic
Figure 1: A simplification of the flow of records through the In Codice Ratio automated system.
ARCHIVING 2021 FINAL PROGRAM AND PROCEEDINGS 85
processing c. 2013. 2016 saw the publishing a report detailing the
opportunities for automation offered by existing eDiscovery sys-
tems,[13] and this work informed the most recent series of experi-
ments which concluded with the publishing of the ‘Using AI for
Digital Records Selection in Government’ report in October 2021.
[14] This report summarized the efforts of five AI vendors to
classify a dataset provided by TNA into retention categories related
to filetype, subject, department, and the like. While paradata coll-
ected by these vendors ultimately informed the conclusions of the
report, it is important to note that in this case the paradata were
collected as artifacts of experimentation and not as a function of
normal operations or operational policy. Such policies have yet to
be developed and would likely change the character and extent of
captured paradata.
Due to the complexity of the dataset and of the requirements
set by TNA, the systems designed and deployed by these five ven-
dors were composed of many different interlinking parts whose
individual tasks were distinct. To closely scrutinize each of these
systems is outside of the scope of this work, but briefly touching
upon the Azure system used by the firm Adatis reveals something
of the bounds of paradata and what opportunities exist for its coll-
ection within complex, multi-part, AI-based automation solutions.
Particularly, this system example highlights how extensive the
paradata may be, as well as the tendency of organizations to already
provide for the collection of some types of paradata in experimental
contexts. In the case of the Azure component of this system, there
exists already fairly comprehensive public documentation, inclu-
ding elements of the source code, information about how the system
handles the trouble-shooting process, design notes, et cetera.[15]
While not every function is explicated, nor is there much discussion
of the nature of changes made between versions, this documentation
evidences the implicit understanding that AI designers already have
regarding what information is necessary to understand the processes
of the systems they have built; even in cases where the final system
includes extensive black-box elements.
Additionally, this example suggests how the construction of a
system might inform paradata collection, especially in cases where
many specialized subsystems are involved. For instance, archivists
may determine that paradata related to the separate subsystems
responsible for data collection are not relevant to a functional under-
standing of the ACS subsystem specifically. (Fig. 2) They may then
opt to treat these different parts of the overall system as having
distinct corresponding paradata. That being the case, strategies for
describing the function of these multi-part systems using paradata
may still have to rely upon a suite of technologies and procedures,
tooled to parse the complex interrelations of component subsystems.
Suggested Paradata Elements
While it is essential to emphasize that the extent of what may
be collected as paradata is still a matter under consideration, the two
aforementioned examples begin to suggest some elements which
should be collected whenever possible for the evaluation of the int-
erpretive processes of AI.
Training Data: This may include full or partial copies of the
training datasets, ideally presented in such a way as to allow
easy examination and interpretation. In cases where training
data uses electronic records already integrated into digital arch-
ival systems, this may be facilitated by the tagging, marking,
and/or linking of these records across fonds.
Performance Information: In addition to quantitative inform-
ation about the embodied strengths and weaknesses of an AI
system, information about performance should also include
that which speaks to the reasons for the existence of particular
confounding variables. This might also include a represent-
ative set of interpretations made by an AI, or information about
the function of diagnostic subsystems.
Versioning Information: Includes proposal and initial design
documentation, analyses of changes made to the structure of
systems or the training strategies upon which they have been
presupposed, and documentation regarding competencies that
develop within a system over time. When possible, this might
Figure 2: A simplification of the flow of records through The National Archives automated processing solution utilizing the Azure Cognitive Search subsystem.
86 SOCIETY FOR IMAGING SCIENCE AND TECHNOLOGY
also include all or part of the source code, and/or maps of the
internal architecture as generated by TensorFlow or similar
platforms for machine learning.
The extension of a particular paradata schema across archives
using the same system or system elements is a potentiality, but it is
important to keep in mind that each archive may differ in their
capacity to collect, store, interpret, and display paradata. Moreover,
guidelines set for a particular system in isolation may not be applic-
able when that system is part of a more complex, networked AI
solution, or in cases where a system can be configured in highly diff-
erentiated ways.
Data Structure for Representation and Display
After determining what and how paradata is to be collected,
archivists will have to determine the means by which paradata may
be organized, interpreted, and displayed. Two current models pro-
vide an opportunity to interrogate how paradata may ultimately app-
ear to different stakeholder groups. Experimentation with these
structures in the archival environment is likely to generate valuable
data about their limitations, leading to improvements in relation to
the needs of archivists and the publics they serve.
Google Model Cards
Originally described in a 2018 journal article by Google resear-
chers,[16] model cards represent a compelling opportunity for the
articulation of collected paradata and conclusions which they have
informed. Model cards are interactive, adaptive modules which
briefly explain the use of an AI-enabled program, provide users with
basic information about its function, and allow users to interact
directly with that program; a working snapshot. Properly configured
model cards can provide valuable general information about AI
capabilities, allow for rudimentary experimentation, and enable
technologically unsophisticated archives users to interact with oth-
erwise inaccessible automated systems.
Current limitations on the usefulness of model cards relate pri-
marily to their reliance on technical experts to encode some of the
modular elements that compose each card. While archives may
outline the varieties of paradata they wish to describe and the
manner in which to describe them, reconfiguration of model cards
is still reliant on a degree of specialized knowledge. Additionally,
model cards are primarily useful in supporting basic, reflexive un-
derstandings of the systems they describe. More complex models,
especially those to be used by engineers and savvy archivists, may
have to rely on different data structures such as IBM AI Fact-Sheets.
IBM AI FactSheets
Launched in 2020, the IBM AI FactSheets project has used the
model of suppliers’ declarations of conformity to experiment with
the collection and arrangement of information relevant to the crea-
tion and deployment of AI modelsinformation which might other-
wise be termed paradata.[17] FactSheets utilize the information
provided by different roles in the AI development cycle (e.g., clients
providing use cases, data scientists providing data gathering strat-
egies) to populate a variety of display templates, each of which are
intended to serve different communications purposes and audiences.
These templates vary in format and complexity from full-scale tech-
nical reports to simplified slideshows, and would theoretically be
generated automatically at different stages in a model’s operational
life cycle.
In comparison to model cards, FactSheets more readily lend
themselves to formal standardization and more plainly comm-
unicate the underlying paradata to the end user. However, Fact-
Sheets are less attractive as a model for displaying paradata to arch-
ives users insofar as they are designed and best-suited to support
professional functions, rather than casual inquiry. Realtime report-
ing via Fact Sheets present an intriguing method to analyze iterative
AI design and could serve to enhance archives governance. How-
ever, the costs of licensing and deploying FactSheets may curtail
their use in more financially-limited contexts.
Directions for Further Research
Using the proposed definition of paradata, future research may
be directed along any number of routes of inquiry. Extensive theo-
retical and practical experimentation must yet be undertaken to
determine the effects that the collection of paradata will have on the
form of AI-based archival automation, and to develop standards.
More information is required about how differing financial,
human, and technological resources will affect the ability of arch-
ives to collect and utilize paradata. It is not yet clear what the
greatest impediments to collecting paradata about AI-based systems
of automation may be; nor is it clear how those obstacles might be
overcome.
Extensive research will need to be conducted into the ability of
archivists to affect the bottom-up development of the AI tools on
which they may soon rely. Further codifying a suite of desirable
features and their complementary paradata would assist in informing
design discussions and ensure that archivists do not find their ethics
have been compromised for the sake of operational expediency.
Research may also be directed towards determining effective
strategies and appropriate forums for the development of paradata
standards and guidelines. Whether the AI-for-archives landscape is
eventually dominated by a single system or remains a site of active
competition, there may be the need for cooperative bodies to create
standards which can be deployed across the archives ecosystem.
Conclusion
In conjunction with increasing interest in the automation of
essential archival functions using artificial intelligence, the time for
the introduction of paradata to the archival sphere has arrived. An
examination of current experiments with artificial intelligence in the
archives suggests that considering and encoding paradata at the
design stage of any AI-driven automation will be essential to satis-
fying the archival ethic. However modest their influence may be,
archives have an opportunity to encourage the development of tools
which are in line with existing archival praxis in pursuit of trans-
parency, accountability, integrity, and usability.
Recent experiments with AI in archives illustrate implicit
understandings of the types of information required to parse the
interpretive products of AI systems and suggest how future project
might iterate with paradata in mind. Existing technologies devel-
oped out of and adjacent to XAI projects represent a compelling
jumping-off point for further experimentation, and extensive oppor-
tunities exist for future research based on the current conception of
paradata and its relation to AI, especially about paradata schema and
standards.
ARCHIVING 2021 FINAL PROGRAM AND PROCEEDINGS 87
While the current InterPARES definition of paradata will likely
be subject to changes, the problems of automation to which it was
created in response will not. Regardless of whether paradata is
ultimately the solution, archivists owe it to the publics they serve to
consider fully their strategies for maintaining archival accountab-
ility in the AI age.
References
[1] M. P. Couper, “Measuring survey quality in a CASIC environment,”
in Proceedings of the Section on Survey Research Methods of the
American Statistical Association, Dallas, TX, USA, 1998, pp. 4149.
[2] H. O’Connor and J. Goodwin, “Paradata,” in Sage Research Methods:
Mixed Methods, Thousand Oaks, CA, USA: SAGE Publications Ltd,
2020.
[3] H. Denard, “The London Charter for the Computer-Based
Visualization of Cultural Heritage,” londoncharter.org, 2009.
[4] A. Bentkowska-Kafel, H. Denard, and D. Baker, Eds., Paradata and
Transparency in Virtual Heritage. Farnham, Surrey, England ;
Burlington, VT: Ashgate, 2016.
[5] I. Huvila, “The Unbearable Complexity of Documenting Intellectual
Processes: Paradata and Virtual Cultural Heritage Visualisation,”
Human IT: Journal for Information Technology Studies as a Human
Science, vol. 12, no. 1, pp. 97110, 2013.
[6] R. Edwards, J. Goodwin, H. O’Connor, and A. Phoenix, Eds.,
Working with paradata, marginalia and fieldnotes: the centrality of
by-products of social research. Cheltenham, UK ; Northhampton,
MA, USA: Edward Elgar Publishing, 2017.
[7] D. A. Wallace, W. M. Duff, R. Saucier, and A. Flinn, Eds., Archives,
Record-Keeping and Social Justice. New York: Routledge, 2020.
[8] R. Guidotti, A. Monreale, S. Ruggieri, F. Turini, F. Giannotti, and D.
Pedreschi, “A Survey of Methods for Explaining Black Box Models,”
ACM Comput. Surv., vol. 51, no. 5, pp. 142, 2019.
[9] Information Innovation Office (I2O), “Broad Agency Announcement
- Explainable Aritificial Intelligence,” Defense Advanced Research
Projects Agency, Arlington, VA, USA, Announcement DARPA-
BAA-16-53, 2016.
[10] S. Ammirati, D. Firmani, M. Maiorino, P. Merialdo, E. Nieddu, and
A. Rossi, “In Codice Ratio: Scalable Transcription of Historical
Handwritten Documents,” in Proceedings of the 25th Italian
Symposium on Advanced Database Systems, Squillace Lido, Italy,
2017, pp. 6572.
[11] E. Nieddu, D. Firmani, P. Merialdo, and M. Maiorino, “In Codice
Ratio: A crowd-enabled solution for low resource machine
transcription of the Vatican Registers,” Information Processing &
Management, vol. 58, no. 5, 2021.
[12] S. Ammirati, D. Firmani, M. Maiorino, P. Merialdo, and E. Nieddu,
“In Codice Ratio: Machine Transcription of Medieval Manuscripts,”
in Digital Libraries: Supporting Open Science, Cham, 2019, pp. 185
192.
[13] The National Archives, “The application of technology- assisted
review to born-digital records transfer, Inquiries and beyond,” The
National Archives, London, UK, Feb. 2016.
[14] The National Archives, “Using AI for digital selection in
government,” The National Archives, London, UK, Oct. 2021.
[15] H. Steen, “Azure Cognitive Search Documentation,” microsoft.com,
Jan. 03, 2022. https://docs.microsoft.com/en-us/azure/search/.
[16] M. Mitchell et al., “Model Cards for Model Reporting,” in
Proceedings of the Conference on Fairness, Accountability, and
Transparency, Atlanta, GA, USA, 2019, pp. 220229.
[17] M. Arnold et al., “FactSheets: Increasing trust in AI services through
supplier’s declarations of conformity,” IBM J. Res. & Dev., vol. 63,
no. 4/5, p. 6:1-6:13, 2019.
Acknowledgements
This paper is an outcome of InterPARES Trust AI, an
international research partnership led by Drs. Luciana Duranti and
Muhammad Abdul-Mageed, University of British Columbia.
InterPARES Trust AI is supported in part by funding from the
Social Sciences and Humanities Research Council of Canada
(SSHRC).
Contributions to this paper made by Jenny Bunn of The
National Archives of the UK are © Crown copyright and are re-
used under the terms of the Open Government Licence.
Author Biographies
Jeremy Davet is a MLIS/MAS student at the School of
Information at the University of British Columbia. Currently he
supports the research efforts of the InterPARES Trust AI as a
Graduate Academic Assistant.
Babak Hamidzadeh has served as the Interim and Associate
Dean of the University of Maryland Libraries; Director of the
RDC at the Library of Congress; in senior management at Boeing;
and in faculty positions at the University of British Columbia and
at the University of Science and Technology in Hong Kong. He
received his Ph.D. in computer science from the University of
Minnesota and is a faculty member at the University of Maryland’s
iSchool.
Dr. Patricia Franks, CA, CRM, IGP, and CIGO is the co-
editor of the Encyclopedia of Archival Science, the Encyclopedia
of Archival Writers, 1515-2015, and the International Directory of
National Archives. She is also author of Records and Information
Management and editor of The Handbook of Archival Practice.
Jenny Bunn is Head of Archives Research at The National
Archives. She has over 25 years’ experience as an archival
practitioner, educator and researcher at institutions including
University College London and the Royal Bank of Scotland. Her
research interests have always lain at the intersection of archives
and technology.
88 SOCIETY FOR IMAGING SCIENCE AND TECHNOLOGY
... It argues that by synthesizing concepts from multiple fields including the empirical social sciences, explainable AI (XAI), and archival studies, paradata provides an approach for supporting the accountability needs of records and recordkeepers in documenting AI processes in specific technical and organizational contexts. As defined by InterPARES for archival contexts, paradata consists of information about the procedure(s) and tools used to create and process information resources, along with information about the persons carrying out those procedures [19]. This paper intends to elaborate on this concept in the AI context specifically; while it is implicit from this definition that paradata exists outside of AI contexts, the essay will limit itself to the direct challenge to transparency of archival processes which emerges in AI implementations. ...
... InterPARES view (shared by this paper) information about the procedure(s) and tools used to create and process information resources, along with information about the persons carrying out those procedures [19] Information about another information resource [66] Common aspects across all sources Information about other information resources; recorded as a means of documenting processes of creation, curation, or management of other information resources Information about other information objects ...
... In this specific context, paradata is a relational category which would fall into the broader category of metadata defined as data about data. However, whereas metadata serves a variety of purposes, paradata records the procedure(s) and tools used to create and process information resources [19]. As Huvila s middle perspective notes, a piece of information which may be understood as paradata at one moment may serve as metadata, provenance, or contextual information from another perspective. ...
Article
Full-text available
The emergence of sophisticated artificial intelligence and machine learning tools poses a challenge to archives and records professionals, who are accustomed to understanding and documenting the activities of human agents rather than the often-opaque processes of sophisticated AI functioning. Preliminary work has proposed the term ‘paradata’ to describe the unique documentation needs which emerge for archivists using AI tools to process records in their collections. For the purposes of archivists working with AI, paradata is here conceptualized as information recorded and preserved about records’ processing with AI tools; it is a category of data which is defined both by its relationship with other datasets and by the documentary purpose which it serves. This paper surveys relevant literatures across three contexts to scope the relevant scholarship which archivists may draw upon to develop appropriate AI documentation practices.
... Sköld et al. (2022) discuss paradata in different information domains as well as close connection and overlapping between concepts of paradata, metadata, and provenance data. Only recently the concept of paradata has been introduced to the archival and recordkeeping sphere in studies focusing on paradata in AI-based automation (Davet et al., 2022(Davet et al., , 2023. As Davet et al. (2023) state, conceptual overlapping exists between the conceptual development of paradata for AI and those of contextual metadata and explainable AI. ...
... When the plans come to flesh in the organization's daily operations, information in the plans becomes paradata about the records. The InterPARES research group defined paradata as "information about the procedure(s) and tools used to create and process information resources, along with information about the persons carrying out those procedures" (Davet et al., 2022). All this information is saved in records management and information control system used. ...
Chapter
Full-text available
To ensure evidence and to control systematically records’ life span in digital environment, we need additional data about records’ background, history, and actions creating those records. National and international specifications typically set requirements for metadata and functionality that an electronic records management system must have. Creation of metadata is resource consuming. One solution to this problem is to hide and automate records management processes. The chapter examines how this has been done in Finnish public administration. Firstly, the chapter contributes to discussion about description of records management processes and adds understanding of possibilities for adding metadata to records. Secondly, we aim to stir up interest toward the use of a concept paradata in recordkeeping and invite discussion of benefits of understanding some of recordkeeping metadata as paradata. While paradata is not an established term in archives and records management, it is a befitting concept to describe information that is gathered about records during their life span.
Chapter
Full-text available
Government provides a variety of critical services, often grounded in decades-old technology. As these services evolve to encompass newer technologies and offer broader capabilities, their exposure to potential threats increases the importance of modeling and simulation to understand their operation and vulnerabilities. This modeling activity requires access to records collected by government agencies, conceptualized formally in terms of archival science. However, the archival science concept of paradata is insufficient to support the forensic reconstruction of the real world required by a modeling effort. This chapter outlines how iterative, interdisciplinary work to model and simulate emergency services communication systems illuminates the need for a new role for paradata in government archives.
Article
Full-text available
Recently introduced into the archival sphere, ‘paradata’ is a conceptual framework for defining the character of information resource processing. (Davet J, Hamidzadeh B, Franks P, Bunn J (2022) Tracking the functions of AI as paradata & pursuing archival accountability. In: Archiving 2022: Final Programs and Proceedings, 7-10 June 2022. Society for imaging science and technology, Springfield, VA, USA, pp 83–88) While it need not be applied exclusively to artificial intelligence-based automated systems, paradata can be, should be, and is currently used to explicate the function of AI in the archives. The use of paradata in relation to AI can help ensure that archival ethical principles continue to be honored in an age of increasing automation. This paper summarizes the theoretical background of paradata, documentation created during recent experiments with machine learning-based AI which gestures toward the varieties of paradata already being collected by researchers, and some of the factors which condition paradata’s use. A few currently available data structures for the representation and display of paradata are evaluated, with an eye toward their suitability for different stakeholder groups. Future directions for theoretical and practical research are suggested.
Article
Full-text available
In the last years many accurate decision support systems have been constructed as black boxes, that is as systems that hide their internal logic to the user. This lack of explanation constitutes both a practical and an ethical issue. The literature reports many approaches aimed at overcoming this crucial weakness sometimes at the cost of scarifying accuracy for interpretability. The applications in which black box decision systems can be used are various, and each approach is typically developed to provide a solution for a specific problem and, as a consequence, delineating explicitly or implicitly its own definition of interpretability and explanation. The aim of this paper is to provide a classification of the main problems addressed in the literature with respect to the notion of explanation and the type of black box system. Given a problem definition, a black box type, and a desired explanation this survey should help the researcher to find the proposals more useful for his own work. The proposed classification of approaches to open black box models should also be useful for putting the many research open questions in perspective.
Article
In Codice Ratio is a research project to study techniques for analyzing the contents of historical documents conserved in the Vatican Apostolic Archives. In this paper, we present our efforts to develop a system to support the automatic transcription of medieval manuscripts, while maintaining the training data collection effort minimal. We focus on crowdsourcing as a means for scalable, expertless training data collection: using crowdsourced character symbols, we train a custom convolutional neural network able to jointly learn correct character shape identification and character recognition. Our approach generates candidate transcriptions by submitting over-segmented character strokes and their combinations to this classifier, while ranking and choosing the most promising outputs by combining the recognition confidence with character and word level statistical language models. We conducted experiments on an unreleased corpus, the Vatican Registers: training our system on 20 pages annotated by the crowd, we were able to obtain good results (19% CER); comparisons to an off-the-shelf system trained with 20 pages annotated with the same process, and to a professional system trained with more than 300 pages transcribed by skilled paleographers demonstrate the opportunities of the proposed approach.
Article
Accuracy is an important concern for suppliers of artificial intelligence (AI) services, but considerations beyond accuracy, such as safety (which includes fairness and explainability), security, and provenance, are also critical elements to engender consumers' trust in a service. Many industries use transparent, standardized, but often not legally required documents called supplier's declarations of conformity (SDoCs) to describe the lineage of a product along with the safety and performance testing it has undergone. SDoCs may be considered multi-dimensional fact sheets that capture and quantify various aspects of the product and its development to make it worthy of consumers' trust. In this paper, inspired by this practice, we propose FactSheets to help increase trust in AI services. We envision such documents to contain purpose, performance, safety, security, and provenance information to be completed by AI service providers for examination by consumers. We suggest a comprehensive set of declaration items tailored to AI in the appendix of the paper.
Chapter
Our project, In Codice Ratio, is an interdisciplinary research initiative for analyzing content of historical documents conserved in the Vatican Secret Archives (VSA). As most of such documents are digitized as images, Machine Transcription is both an enabler to the application of Knowledge Discovery techniques, as well as a useful tool to the paleographer for speeding up the transcription process. Our approach involves a convolutional neural network to recognize characters, statistical language models to compose and rank word transcriptions, and crowdsourcing for scalable training data collection. We have conducted experiments on pages from the medieval manuscript collection known as the Vatican Registers. Our results show that almost all the considered words can be transcribed without significant spelling errors.
Conference Paper
Trained machine learning models are increasingly used to perform high-impact tasks in areas such as law enforcement, medicine, education, and employment. In order to clarify the intended use cases of machine learning models and minimize their usage in contexts for which they are not well suited, we recommend that released models be accompanied by documentation detailing their performance characteristics. In this paper, we propose a framework that we call model cards, to encourage such transparent model reporting. Model cards are short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions, such as across different cultural, demographic, or phenotypic groups (e.g., race, geographic location, sex, Fitzpatrick skin type [15]) and intersectional groups (e.g., age and race, or sex and Fitzpatrick skin type) that are relevant to the intended application domains. Model cards also disclose the context in which models are intended to be used, details of the performance evaluation procedures, and other relevant information. While we focus primarily on human-centered machine learning models in the application fields of computer vision and natural language processing, this framework can be used to document any trained machine learning model. To solidify the concept, we provide cards for two supervised models: One trained to detect smiling faces in images, and one trained to detect toxic comments in text. We propose model cards as a step towards the responsible democratization of machine learning and related artificial intelligence technology, increasing transparency into how well artificial intelligence technology works. We hope this work encourages those releasing trained machine learning models to accompany model releases with similar detailed evaluation numbers and other relevant documentation.
Measuring survey quality in a CASIC environment
  • M P Couper
M. P. Couper, "Measuring survey quality in a CASIC environment," in Proceedings of the Section on Survey Research Methods of the American Statistical Association, Dallas, TX, USA, 1998, pp. 41-49.
The London Charter for the Computer-Based Visualization of Cultural Heritage
  • H Denard
H. Denard, "The London Charter for the Computer-Based Visualization of Cultural Heritage," londoncharter.org, 2009.
Working with paradata, marginalia and fieldnotes: the centrality of by-products of social research
  • R Edwards
  • J Goodwin
  • H O'connor
  • A Phoenix
R. Edwards, J. Goodwin, H. O'Connor, and A. Phoenix, Eds., Working with paradata, marginalia and fieldnotes: the centrality of by-products of social research. Cheltenham, UK ; Northhampton, MA, USA: Edward Elgar Publishing, 2017.
In Codice Ratio: Scalable Transcription of Historical Handwritten Documents
  • S Ammirati
  • D Firmani
  • M Maiorino
  • P Merialdo
  • E Nieddu
  • A Rossi
S. Ammirati, D. Firmani, M. Maiorino, P. Merialdo, E. Nieddu, and A. Rossi, "In Codice Ratio: Scalable Transcription of Historical Handwritten Documents," in Proceedings of the 25th Italian Symposium on Advanced Database Systems, Squillace Lido, Italy, 2017, pp. 65-72.