Introduction
The idea to use graphs to represent knowledge, which can be automatically
actioned upon by machines, has been around since the early 60’s. Both, Engelbart
[
9
] (in 1962) and Lickleder [
18, 23
] (in 1965) imagined a future where machines
would be able to automatically process and reason over data represented in
knowledge graphs. Almost forty years later the seminal Semantic Web paper by
Berners-Lee et al. [
3
] described their vision of a Semantic Web, whereby the
existing web infrastructure could be used to represent data in a manner that
could be automatically actioned upon by intelligent software agents.
Roughly five years after the seminal paper both Shadbolt et al. [
25
] and
Feigenbaum et al. [
10
] reflected on the state of the art at the time and concluded
that although intelligent agents were still far from being realised the technology
was steadily gaining traction especially as a means of data integration. More
recently, Glimm and Stuckenschmidt [
13
] and Bernstein et al. [
4
] confirm, that
approximately 17 years on, the vision has yet to become a reality. The authors
observe that although the primary focus of the Semantic Web community was
initially on knowledge representation, reasoning and querying, in recent years there
has been a broadening beyond pure Semantic Web topics to include knowledge
extraction, discovery, search and retrieval. However, according to Bernstein et al.
[
4
] there are still a number of issues concerning heterogeneity both in terms of
representation and semantics, and diversity in terms of web data quality.
Additionally, the authors identify new challenges that arise with increasing data
volume and publishing velocity.
Although Bernstein et al. [
4
] point to several challenges that the
community will face in the future, there is a distinct lack of focus on topics that are
important from an intelligent agents perspective, yet remain under represented
within the community. For instance, technologies, techniques and protocols that
enable agents to interact with other agents in order to carry out their activities
according to constraints in the form of goals, preferences, and usage limitations
in a trustworthy manner.
In order to fill this gap, this position paper revisits the original vision of
the Semantic Web and highlights several open research questions that are
important if we hope to one day have intelligent agents that are able to act on
our behalf. We start by examining the status quo in terms of existing Linked
Data management practices. In particular, we discuss data management through
the lens of the FAIR3 (Findable, Accessible, Interoperable and Reusable) data
principles, and highlight several open research challenges in terms of persistent
identifiers, indexing, and usage constraints. Following on from this we argue
for adapting and extending these FAIR principles to guide the development of
FAIR ICT Agents, whereby ICT denotes Interactive intelligent agents that are
Constrained via goals, preferences, norms and usage restrictions, in a manner
that fosters Trustworthiness.
The remainder of the paper is structured as follows: Section 2 introduces the
FAIR data principles and discusses some of the limitations of current Linked
Data management practices. Section 3 presents several challenges and
opportunities that need to be overcome for FAIR ICT Agents to become a reality. Finally
Section 4 concludes the paper and identifies several open research questions.
2
Making Linked Data FAIR
The Semantic Web enables things, otherwise known as resources represented
using the Resource Description Framework (RDF) data model, to be linked
using Internationalised Resource Identifiers (IRIs) in a similar way to how web
documents are linked using the HyperText Markup Language (HTML)
hypertext reference (HREF) attribute. Linked Data is a related concept, which refers
to a set of best practices for publishing and connecting structured data on the
Web [
19
]. In recent years, we have seen significant advances in the technology
used to both publish and consume Linked Data, however a recent article by
Beek et al. [
2
] claims that the existing Semantic Web is neither traversable nor
machine-processable, and consequently argues that the Semantic Web needs
centralisation. In this position paper, we argue for treating the root cause of the
problem (i.e., highlighting existing data management challenges and calling for
best practices guidelines and research to address them) rather than the
symptoms (i.e., developing centralised solutions on top of distributed web data that
3 FAIR data principles, https://www.force11.org/node/6062
address the inherit limitations of the existing infrastructure). In terms of the
former we argue that a necessary first step is to provide additional guidelines for
data publishers that go beyond the original Linked Data principles and the well
known 5 star rating system4.
An emerging best practice in terms of scientific knowledge dissemination is
the adoption of FAIR data principles [
29
], whereby researchers strive to
ensure that their research objects (papers, datasets, code etc...) are Findable,
Accessible, Interoperable and Reusable. Although the FAIR principles were
devised to provide guidance for managing scholarly assets, we believe that said
principles could be adapted to provide guidance to Linked Data publishers in
order to improve the findablility, accessibility, interoperability and reusability of
machine readable data available on the Web.
2.1
FAIR data
The core objective of the FAIR data principles is to provide guidance to scholarly
data publishers in terms of making their data reusable by both humans and
machines. The four foundational principles can be summarised as follows:
– To be deemed Findable, data should be uniquely identifiable via persistent
identifiers, these identifiers should be used to associate descriptive metadata
with the data, and both data and metadata should be indexed in a manner
that is easy to search.
– In order to make data Accessible it should be possible to retrieve the data
via common protocol(s), that are open, free, universally implementable and
can support usage constraints where desirable.
– Making (meta)data5 Interoperable is primarily concerned with the
representation of (meta)data in a manner that facilitates integration e.g. using
common/standard ontologies and vocabularies.
– Finally, (meta)data is Reusable if it is richly described in terms of relevant
attributes, contains relevant provenance information and is compatible with
domain specific standards.
2.2
FAIR Linked Data
There is clearly a strong connection between said principles and Semantic Web
technologies and Linked Data principles. Both Reusability and Interoperablity
are at the core of the Resource Description Framework (RDF) data model. By
using RDF to describe resources, it is possible to describe complex relations
between resources in a machine readable format. Ontologies provide for a shared
understanding of things and how they are related, that can easily be reused
and extended. Data is linked to other data using HyperText Transfer Protocol
(HTTP) IRIs that can be used to identify things (papers, datasets, code etc...).
4 https://www.w3.org/DesignIssues/LinkedData.html
5 In order to improve readability in this paper we use (meta)data to denote to data
and metadata.
Although the RDF data model and Linked Data principles are good starting
points in terms of making Linked data FAIR, there are still a number of open
research challenges. In terms of Findablility, according to FAIR (meta)data
should be identifiable via persistent identifiers . Despite a push by the
community to use persistent identifiers, for instance for resources submitted
to the International Semantic Web Conference (ISWC) resources track6, they
are still not widely used in practice. Another key aspect of Findablility is the
indexing of (meta)data in a manner that is easy to search. Although
there have been a number of proposals (cf. [
2, 11
]), given that indexing is done in
a centralised manner existing proposals sufer from data freshness issues. From an
Accessibility perspective when it comes to usage constraints that describe
how the data should be used there is a large body of work on access control
specicfiation and enforcement strategies for RDF [
21
] and licensing [
14, 15, 16, 27
]
proposals for data exposed as Linked Data (cf. Section 3 for additional details).
The challenge here is the fact that existing usage control strategies (where used)
are still very primitive.
Although FAIR was devised to provide guidance in terms of efective
scholarly data management, we argue that by adapting the FAIR data principles for
Linked Data it will be possible not only to identify existing challenges in terms of
data management, which we only touch upon in this article, but also to provide
a best practice guide for dealing with these challenges.
3
Towards Intelligent Agents
Returning to the original visions by Berners-Lee et al. [
3
], whereby intelligent
agents are able to make sense of Web data and alleviate humans from performing
complex analytical tasks, it is clear that FAIR principles alone are not enough
as they focus on data management without considering how intelligent agents
might make use of this data. In this respect, we identify the need for FAIR ICT
Agents, whereby ICT denotes Interactive intelligent agents that are Constrained
via goals, preferences, norms and usage restrictions, in a manner that fosters
Trustworthiness.
3.1
Interactive intelligent agents
When it comes to intelligent agents the services ofered by each agent need to
be designed in a manner such that multiple agents can interact (and possibly
even collaborate) in order to complete tasks and solve problems. Each agent
needs to maintain a list of services that it is capable of executing based on the
(meta)data in its knowledge graph (including descriptive attributes, constraints
and provenance data). Ideally, the list of services should grow organically with
the data and as the agent uncovers new insights based on incremental analysis
of its knowledge graph.
6 http://iswc2018.semanticweb.org/call-for-resources-track-papers/
Unlike traditional web services, semantic web services use formal
ontologybased annotations to describe the service in a manner that can be automatically
interpreted by machines. In the early years of the Semantic Web there were
several standardisation initiatives, namely the Web Ontology Language for Web
Services (OWL-S)7, the Web Service Modeling Language (WSML)8, the W3C
standard Semantic Annotations for WSDL and XML Schema (SAWSDL)9. A
survey conducted by Klusch et al. [
22
] provides a summary of existing work and
describes the various semantic web service search architectures (i.e. centralised
and decentralised directory based, and decentralised directoryless). The authors
conclude that research into decentalised semantic service search is
lagging far behind its centralised counterpart. When it comes to semantic
web services the big question is how do we support adaptive discovery and
composition of semantic services? Other open research challenges are
concerned with enabling interoperability between policy aware agents, and
dealing with agents joining and leaving the network at will.
3.2
Constrained via goals, preferences, norms and usage restriction
Berners-Lee et al. [
3
] originally envisioned a system, where intelligent agents
were capable of acting on behalf of humans. One of the key components of such
a system is the policy language that is capable of capturing the constraints
under which the agents operate. During the early days of the Semantic Web
the development of general policy languages that leverage semantic technologies
(such as KAoS [
7
], Rei [
20
] and Protune [
6
]), was an active area of research.
General policy languages cater for a diverse range of functional requirements
(e.g., access control, query answering, service discovery, negotiation, to name
but a few). Considering that the policy language needs to be interpreted by
machines, formal semantics is important as it allows for the verification
of correctness. However, research into general semantic policy languages seems
to have reduced considerably in recent years and the suitability of existing
general policy languages towards the intelligent agents vision is still an
open research question.
In terms of specific policy languages access control is a topic that has
received a lot of attention over the years. Kirrane et al. [
21
] provide a detailed
survey of the various access control models, standards and policy languages,
and the diferent access control enforcement strategies for RDF. Although there
have been several diferent proposals over the years, there is still no standard
access control strategy for Linked Data. Considering the array of access
control specification and enforcement mechanisms proposed to date, a necessary
ifrst step towards ensuring that intelligent agents have the ability to decide
with whom they share information is to develop a framework that can be
used to evaluate existing access control oferings in terms of
expressivity, correctness and completeness. When it comes to usage control in the form
7 https://www.w3.org/Submission/OWL-S/
8 https://www.w3.org/Submission/WSML/
9 https://www.w3.org/TR/sawsdl/
of licensing, research topics range from using Natural Language Processing to
extract license rights and obligations [
8
] to licenses compatibility validation and
composition [
14, 15, 16, 27
]. More recently, the Open Digital Rights Language
(ODRL)10, which became a W3C recommendation in February 2018, provides
a promising first step towards the general adoption of machine understandable
licenses, however it remains to be seen if data publishers embrace the
new standard and if license aware data querying and processing
mechanisms become common practice.
Another important research direction that remains underdeveloped is the use
of policies to specify societal norms and personal values that would
enable agents to understand the constraints of the environment in which
they operate. Also, there are also several open research questions in terms of
the suitability of the existing languages to deal with the volume, velocity,
variety and veracity of data we are faced with today, the ability to
balance expressivity and computational complexity, and ensuring that the
intelligent agent ecosystem can deal with the policy interoperabiliy needs of
collaborating agents.
3.3
Fostering trustworthiness
Artz and Gil [
1
] conducted a comprehensive survey of trust mechanisms in
computer science in general and the Semantic Web in particular. The authors
highlight that traditional approaches focused primarily on authentication via
assertions by third parties, however in later years the topic evolved to include
historical interaction data, the transfer of trust from trusted entities, and
decentralised trust mechanisms (e.g. voting mechanisms or other consensus decision
making mechanisms). Although there is a large body of computer science
literature relating to trust the efectiveness of existing trust mechanisms in
the context of intelligent agents has yet to be determined.
In an intelligent agent ecosystem local provenance chains could be used by
agents to provide explanations for decisions made, while global provenance chains
could be used to provide transparency with respect to collaborating agents or
the distributed system as a whole. These provenance chains could also be used to
record and retrieve historical data and to build trust between agents. Although
there has been a number of proposals for representing provenance events (cf.
[
12, 17
]). To date the focus has been on recording where the data came from or
capturing the source of the data or changes to data over time. In this regard there
have been several standardisation initiatives, such as PROV 11 and OWL-Time12
ontologies, that can be used to represent provenance and temporal information
respectively. In the context of intelligent agents there is a need to record
provenance with respect to both data and processing in a manner
that can be easily digestible.
10 https://www.w3.org/TR/odrl-model/
11 PROV,https://www.w3.org/TR/prov-overview/
12 OWL-Time,https://www.w3.org/TR/owl-time/
From a provenance chains perspective there are two distinct avenues that
could be leveraged, one built on top of existing web protocols [
24, 28
] and another
based on blockchain technologies [
30
]. Weitzner et al. [
28
] present their vision of
a policy-aware architecture for the Web, which includes three basic components:
policy-aware audit logging, a policy language framework, and accountability
reasoning tools. Specicfially, they discuss how transparency and accountability can
be achieved via distributed accountability appliances that communicate using
existing web protocols. Seneviratne and Kagal [
24
] build on this idea by proposing
a distributed accountability platform known as Accountable Hyper Text
Transfer Protocol (HTTPA) that allows data producers to express usage restrictions
and data consumers to express usage intentions. Unfortunately the authors only
touch upon the required features and the proposed accountability platform
has yet to be assessed from both a functional or a non-functional
requirements perspective. Alternative distributed architectures for transparent
personal data processing are discussed by Bonatti et al. [
5
], however the authors
simply describe the opportunities and challenges, and the concrete
implementation is left to future work. Zyskind et al. [
30
] discuss how blockchain technology
could be extended to keep track of both data and access transactions. One of
the primary drawbacks of the work is the fact that the authors focus on how
to repurpose the blockchain as an access-control moderator as opposed to
exploring the suitability of the proposed architecture for data transparency and
governance. Another related avenue of research by Third and Domingue [
26
]
proposes a semantic index for distributed ledgers.
Although, Blockchain platforms such as Ethereum13 and Hyperledger
Fabric14 have the capability to support policy aware service provision, via smart
contracts and chaincode, the suitability of blockchain platforms in terms
of both functional and non functional requirements remains an open
research question. In addition, there are a variety of societal challenges that
also need to be considered, such as the right to be forgotten, algorithmic
biases, fake news, filter bubbles , to name but a few.
4
Conclusion and Future Work
In this paper, we revisit the original vision of the Semantic Web whereby software
agents are able to perform complex computational tasks on behalf of humans
[
3
]. Inspired by recent surveys [
4, 10, 13, 25
] that analyse the evolution of
Semantic Web technologies over almost two decades, we strive to shed light on
important research topics that are necessary for the development of intelligent
agents however are currently under represented at the predominant Semantic
Web publishing venues. In order to frame the discussion we started by
examining existing Linked Data publishing and consumption practices through the
lens of the scientific FAIR data principles. From a data management perspective,
we identified several open research challenges in terms of persistent identifiers,
13 https://www.ethereum.org/
14 https://www.hyperledger.org/projects/fabric
indexing, and usage constraints. From an application perspective, we argued for
additional research to support interaction between agents that are constrained
via goals, preferences, norms and usage restrictions, in a manner that fosters
trustworthiness in the services delivered. From a best practices perspective, a
potential first step is to adapt and extend the FAIR data principles such that
they can serve as a best practice guide for FAIR ICT Agents. We do not claim
that this is an exhaustive list of challenges, but rather with this position paper
we hope to rejuvenate interest in these under represented topics with a view to
bringing us closer to making the intelligent agent vision a reality.
Acknowledgments. Supported by the Austrian Federal Ministry of Transport,
Innovation and Technology (BMVIT) DALICC project https://www.dalicc.net.