In this paper we outline the main lines of research for de ning a framework based on Linked Open Data (LOD) for supporting knowledge creation in the Cultural Heritage (CH) eld with a particular focus on History of Art research. We delineate the main challenges we need to deal with and we explore the state-of-the-art in LOD publishing systems, LOD citation and authority management. Furthermore, we introduce the idea of computer-aided serendipity in History of Art research with the purpose of contributing to the advancement of the eld and to the de nition of new methodologies for entity linking and retrieval.
One of the most relevant socio-economical and scienti c changes in recent years
has been the recognition of data as a valuable asset. The Economist magazine
recently wrote that \data is the new raw material of business"; and the European
Commission stated that data-related \technology and services are expected to
grow from EUR 2.4 billion in 2010 to EUR 12.7 billion in 2015" [
The CH domain and in particular the History of Art research eld, provide a fertile ground where the LOD potential can grow and bloom; indeed, in History of Art the preponderant way to produce new knowledge is to reveal connections between di erent items (illuminated manuscripts, pictures, frescos) that can cast a new light on an artist, an artistic movement or an art-historical period.
This potential is widely recognized by public and private agencies, which
keep investing in publishing CH resources as LOD. In the CH domain, which in
the EU alone \accounts for 3.3% of GDP and employs 6.7 million people (3% of
total employment)" [
Nevertheless, publishing CH resources as LOD on the Web, is just the rst step to enable new knowledge creation. To realize the full potential of the Web of Data, we need to devise innovative methods for seeking and creating links between datasets and to design user-oriented services for exploiting them. To this purpose, History of Art is a fruitful domain that, not only can bene t, but also can assist the development of these new methods because it provides: rich and heterogeneous information needs, ample structured and unstructured resource datasets and a proactive research community accustomed to seek new connections between entities.
The aim of this paper is to outline the basics to design a LOD framework for supporting knowledge creation in the History of Art domain and to outline some of the scienti c challenges we have to face. Overall, the framework we envision pursues two main objectives: (i) to provide LOD-ready functionalities to create, access and cite new knowledge and to represent and retain authority information; (ii) to assist domain experts in the creation of semantic connections and to empower them with (semi) automatic serendipity capabilities.
Accomplishing the rst goal will overcome current LOD publishing practice where the data are produced and stored by local systems and then mapped, synchronized and published as LOD; this procedure requires a multiplication of investments and produces scarcely connected datasets given that the expert knowledge is not tackled in the publishing activity. Our idea is to con ate data creation, connection, and publishing in one phase and let computer scientists and domain experts to work back-to-back joining knowledge and e orts. In this context there is the need also to provide concrete solutions for fundamental but overlooked issues as data citation and authority management.
Achieving the second goal will push History of Art research boundaries by joining the experts' knowledge and intuition with automatic tools able to retrieve relevant entities in the Web of Data. To this purpose, from the computer science point-of-view, we need to devise query models from the experts information needs, envisage and de ne methods for mapping entities to queries, use them for retrieving target entities and then provide entity rankings. Current solutions are biased toward Web search where user needs are expressed using very few terms; as a consequence, they employ at models of user inputs where contextual information, text, images, links, categories and feedbacks, when available, are considered at the same level and equivalent one to the other.
The History of Art domain provides us with rich and heterogeneous user inputs that poses additional challenges, but that also provides the means for advancing entity retrieval state-of-the-art. The framework we envision could also
push ahead the creation and distribution of CH resources o ering new ways for addressing consumers' demand for data access and for greater participation in the creativity process.
The rest of the paper is organized as follows: Section 2 presents the state-ofthe-art in LOD publishing in the CH context, serendipity oriented algorithm and data citation methodologies. In Section 3 we present the main design lines, the challenges we have to face for realising the envisioned framework and introduce the art research use case on which the proposed framework will rely. In Section 4 we propose an architecture that could implement the envisioned framework. Finally, in Section 5 we draw some nal remarks. 2
The e orts for disclosing the LOD potential within the CH eld must go towards the design of new methodologies for creating meaningful and possibly unexpected semantic links between data and for managing the knowledge created through these connections. This endeavor is needed to make the Web of Data fully operational and so that its role can be compared to that played by search engines over the last two decades in disclosing the full potential of the Web.
The CH domain and particularly the History of Art not only can be bene cial of the full exploitation of the LOD potential, but they can also provide fertile ground where new methods and technologies can be designed, developed and applied. Indeed, in History of Art the preponderant way to produce new knowledge is to reveal connections between di erent items (illuminations, pictures, frescos) that can cast new light on an artist, an artistic movement or an art-historical period. The most valuable connections are the unexpected ones linking elements that may seem to have very few in common; indeed, in this domain, important discoveries have often been done thanks to associations and connections of items emerged also by chance in the research path identi ed by domain experts. This process of discovery can be de ned as serendipity and it is especially encouraged by LOD where meaningful links between entities allow us to move across diverse and apparently unrelated knowledge domains. History of Art is a well suited domain where algorithms fostering serendipity within LOD can be designed, developed and evaluated because it provides: rich and heterogeneous information needs, ample structured and unstructured resource datasets and a proactive community accustomed to seek new semantic connections between entities.
Given the scienti c and economic impact of LOD and the leading role of CH
for triggering its potential, a lot of research is being carried out in important elds
such as database systems, information retrieval and digital libraries focusing
especially on e cient ways for storing and querying datasets [
However, several crucial questions still remain unanswered or overlooked. What merits to be linked? How can linking towards potentially unknown data sources be eased? How can the serendipity process be enhanced and guided by semantic methodologies? How can domain experts' knowledge be exploited for creating meaningful semantic links? How the authoritative information about entities, activities and people involved in data creation be established and retained? How new created knowledge represented by a data subset can be cited?
We can point out four main research areas concerning these questions that
must be taken into account:
{ LOD-based systems: The vast majority of systems in the CH domain
are not natively LOD-based and adopt a publishing paradigm involving
redundancy and multiplication of costs [15]. Moreover, in most cases (e.g.
Europeana), the links to external sources are established without involving
domain experts and are rather syntactic than semantic. The framework we
envision is aimed as developing user-oriented services for native LOD
creation and exploration. Moreover, it aims to provide services for connecting
data while experts are working, thus exploiting their knowledge for link
creation.
{ Serendipity capabilities: No system in the CH domain provides end-users
with this function. Related methodologies are proposed in the context of
entity linking and retrieval targeting Web search: they deal with sparse user
inputs [
For what it is concerned with the serendipity algorithms the framework we envision will employ (semi) automatic methods for suggesting entities to researchers; the idea is that this methodology could help them in their daily work by triggering the exploration of new research paths leading to new knowledge creation in History of Art. Now, we present the concrete steps we are going to follow to
realize the framework that we call Linked Open Data Enhanced Framework for Serendipity in Art Research (LODESTAR).
We will rely on a concrete use case rooted in history of art research which aim is to identify, analyze and assess the in uence of the most ancient Divine Comedy illuminated manuscript illustrations on the many representations of the Hereafter and biblical and mythological characters that can be found in lateXIV century and Renaissance art. Indeed, many inventions that can be found in the rst Divine Comedy illuminated manuscripts were destined to have great success and also to be used in later manuscripts, not only with Dante's poem, but also with other texts and other works of art. In order to achieve large-scale results, beside illuminations, we will also take into account products of the major arts such as paintings, frescos and sculptures, thus bringing new knowledge on di erent elds of the History of Art. The object of the research is highly original and constitutes a eld of investigation that has never been explored before. Given its ample scope, the large di usion of the Divine Comedy and its in uence through the centuries on many elds, this use-case o ers substantial chances of success especially for applying semi-automatic serendipity methodologies; it is a valuable starting stage to design, apply and test methods that will be then re-used in other use-cases and domains.
Input streams extracted from the environment
LODESTAR work environment
Rich and heterogeneous inputs
related images multimodal related documents
item connections domain experts descriptions
In Figure 1 we show the work ow realizing the objectives tackled by the LODESTAR framework. In the center there is the work environment where domain experts will carry out their daily research work. This environment provides rich and heterogeneous input streams such as free and semi-structured text, images, links, user pro le and contextual information; LODESTAR is required to automatically catch these streams that are implicit expressions of user needs and to extract entities from them.
These entities will be sources for the entity retrieval algorithms realizing the LODESTAR serendipity methodology shown on the right hand side of Figure 1. This would be a major shift from state-of-the art entity retrieval solutions, which are typically employed in syntactic linking processes where input sources are constituted by few keywords. In LODESTAR users are not asked to explicitly issue queries, but the traces they leave behind while using the system will be automatically gathered and exploited for devising queries. To tackle this goal we have to employ new methods to model the extracted entities by associating di erent groups of entities to di erent levels of relevance decided on the basis of the user needs within the context under consideration; in other words we have to go beyond the current \ at models" that consider all the user inputs at the same level.
This model will allow us to devise several queries corresponding to di erent entities selections and compositions that will be used to retrieve \target entities" from the LOD cloud. The purpose of issuing multiple queries is to cover a wide spectrum of potential user needs as well as to ease the discovery of unknown entities that may trigger the creation of unexpected connections. The rank module gathers the target entities and orders them accordingly to the user need they tackle.
Lastly, the ranked entities are suggested to the users that will consider them for establishing semantic connections fostering new knowledge creation.
The left-hand side of Figure 1 reports data citation and authority
management components needed to give credit to data creators (e.g. produce a
diachronic reference to a data subset), retain context (e.g. who created a link,
when a link has been created) and control data con icts (e.g. choose the most
authoritative statement among contradicting ones). LODESTAR has to provide
general methodologies for tackling these issues for which, in literature, there are
no concrete solutions yet. We will employ the solution proposed in [18] by
extending it in order to consider also the temporal aspects of LOD datasets; indeed,
a citation should be resolved by referring to the precise version of the data that
was cited. To this end we plan to employ some RDF versioning methodology as
described in [
Serendipity can be realized by following the work ow shown on the right of Figure 1. Firstly, we need to trace the user information within the LODESTAR work environment such as the item being studied, annotations, items being compared to it, free text and semi-structured text (e.g. form elds), clickstream and user pro les. Then, we apply mining algorithms, extending existing entity extraction solutions, for extracting phrases, entities, categories from all these input streams; afterwards, the enhancement module, exploiting suitable ontologies, will connect the extracted entities one with the other via automatically established semantic links. At this stage we obtain an RDF graph of connected entities representing a very rich combination of user needs.
There exists no ready-to-use solution for handling this \user needs graph" and devise queries from it; LODESTAR represents the very rst e ort tackling this problem. The methodology we envision is to hierarchically model the user needs graph by associating di erent groups of entities to di erent levels of relevance decided on the basis of contextual information.
Machine learning methods have been proposed in literature for the task of
learn classes in ontologies and constructing knowledge from user-provided
examples; a possibility is to develop a similar approach to assign relevance weights
or probabilities (i.e. the estimated relatedness to the user needs) to the entities
in the graph. These weights will be exploited for modeling the user needs graph
with a innovative set-based model, i.e. the NESTOR model [
I
B
A C
D
S+(B) = {X 2 I | X ⇢ B} = {C, D} E
BE = B [ E B = B \ S+(B) B
C
D B
C
D
E
qS+(B) qBE qB
NESTOR comes with a whole bunch of de ned set-operations that will be exploited to design an algorithm for selecting alternative entity combinations within the graph; from these selections a bunch of queries covering a wide spectrum of potential user needs will be devised.
The queries will be issued against the LOD cloud for retrieving groups of target entities; each group will be ordered by a ranking model built on solutions for structured search, e.g. statistical language models, and then aggregated by exploiting meta-search engine techniques to be presented as link suggestions to the users.
This envisioned methodology can be embedded in a more comprehensive architecture as shown in Figure 3 composed by 4-staked layers concerning important aspects of a system architecture: scalability, robustness, reactivity and suitability.
The rst layer is designed by considering that LODESTAR will deal with big datasets and large user and machine requests. Therefore, in order to provide linear scalability and fault tolerance, several technologies should be taken into L D E S t A R
framework y t i l i b a t i u S Analysis
Service
CREATE y itv SPARQL itc end-point a e R s s e tn Data su Integration b o R y t i l i b a l a c S
Domain Expert Community Search Service FIND
Visualization
Service
Linking Service
Citation
Service EXPLORE
INNOVATE
ATTRIBUTE
Entity Extraction and Enhancement
User's Inputs
Modeling
Adaptive
Querying
SERENDIPITY CAPABILITIES Ontology
Data Citation
Authority Management Map-Reduce Programming
Model
Hadoop Distributed File
System
Mapping and
Synchronization
Distributed Multinode Cluster
Fig. 3. A possible architecture implementing the LODESTAR framework. account, in particular: Apache Cassandra4 as distributed column store solution and Apache Hadoop5 in conjunction with the Map-Reduce functionalities to provide reliable and distributed computing.
Scalability is also concerned with the selection of datasets that will serve the use-case we have in mind; in a real-world scenario we cannot assume that all relevant data will be available as LOD because even if many relevant CH institutions and systems are LOD-compliant { e.g. Europeana6, the Library of Congress, the British Library, the French National Library, and Cultura Italia7 { others still are not { e.g. Artstor, Dante On-line8, and the Web gallery of Art.
To deal with all these sources, it is necessary to employ a mapping and synchronizing module that will harvest, map to LOD and store these data. An apposite data integration module will blend di erent resource kinds (as part of the robustness layer). All the resources will be represented accordingly to ontologies de ned within the domain of interest; computer scientist and art historians will jointly carry out this activity by de ning an RDF Schema and reusing existing vocabularies to establish bridges with third-party ontologies { e.g SKOS, Europeana Data Model, CIDOC-CRM { that will also be exploited for serendipity.
In the reactivity layer, we could provide a SPARQL end-point, which will work as a data provider allowing third-party services to discover, re-use and connect to the LODESTAR data. This layer also accommodates the modules implementing the serendipity algorithms.
All the presented services and solutions must be made available to end-users via pluggable user interface software components (i.e. Web portlets); the main end-user interfaces need to deal with an environment for describing items, annotating and visually comparing them, seeking and connecting entities, browsing the data and making use of the data citation and serendipity modules. 5
In this paper we discussed the role of LOD in the CH eld and we outlined some of the cultural and computational challenges we have to face in order to de ne a framework for supporting knowledge creation in CH. In particular, we selected a stimulating use case based on History of Art research which provides us with rich input data and a research community oriented to seek new connections between di erent items to create new knowledge.
The main challenges we outlined regard the necessity to de ne new methodologies for entity linking and retrieval which take into account complex user inputs, a work environment which on the one hand exploits existing LOD for knowledge creation and on the other hand stimulates the production of new and 4 http://cassandra.apache.org/ 5 https://hadoop.apache.org/ 6 http://www.europeana.eu/ 7 http://www.culturaitalia.it/ 8 http://www.danteonline.it/ unexpected connections between CH resources and a data citation methodology allowing us to cite evolving subset of data with variable granularity and to produce both machine- and human-readable references. 15. Marden, J., Li-Madeo, C., Whysel, N., Edelstein, J.: Linked Open Data for Cultural Heritage: Evolution of an Information Technology. In: Proc. of the 31st ACM International Conference on Design of Communication. pp. 107{112. SIGDOC '13, ACM, New York, NY, USA (2013) 16. Meij, E., Bron, M., Hollink, L., Huurnink, B., de Rijke, M.: Mapping Queries to the Linking Open Data Cloud: A Case Study Using DBpedia. Web Semant. 9(4), 418{433 (2011) 17. Proll, S., Rauber, A.: Scalable Data Citation in Dynamic, Large Databases: Model and Reference Implementation. In: Proc. of the 2013 IEEE International Conference on Big Data, 6-9 October 2013, Santa Clara, CA, USA. pp. 307{312 (2013) 18. Silvello, G.: A Methodology for Citing Linked Open Data Subsets. D-Lib Magazine 21(1/2) (2015), http://dx.doi.org/10.1045/january2015-silvello 19. Stefanidis, K., Chrysakis, I., Flouris, G.: On Designing Archiving Policies for Evolving RDF Datasets on the Web. Lecture Notes in Computer Science, vol. 8824, pp. 43{56. Springer (2014) 20. W3C: RDF 1.1 Concepts and Abstract Syntax { W3C Recommendation 25 February 2014 (February 2014)