In recent years, the Resource Description Framework (RDF) has become the de-facto standard to represent heterogeneous semi-structured data on the web. RDF datasets are interrogated with SPARQL, a structured query language which is often not intuitive for the nonexpert users, due to its syntax and the necessity to know the structure of the underlying graph. A simpler paradigm like keyword search can help in this regard to access these databases. Moreover, nowadays datasets constitute the backbone of the scienti c research, and thus they should be cited as any other scholarly publication. RDF presents a new challenge in the automatic creation of textual citation since it lacks the structure of RDB and XML databases. In this work, we discuss the design and development of a system which will perform keyword-search on RDF graphs and, given the results, will create the textual citation for the nal user.
In this paper, we describe the general structure of a system for the extraction of data through keyword search from an RDF database and the automatic creation of a human-readable citation snippet from these data.
Keyword Search In the recent years, RDF, a family of W3C speci cations
for the creation of directed graph databases, has become the de-facto standard
for the publication, the access and the sharing of data on the Web. This because it
allows for exible manipulation, enrichment, discovery, and reuse of data across
applications, enterprises, and community boundaries. Recently, the growth of
large knowledge-sharing communities like Wikipedia and the advances in the
automated information extraction from Web pages have enabled the creation of
large-scale knowledge bases [
RDF graphs can be interrogated through the SPARQL structured language. This language is di cult for non-expert users due to its complex syntax and the necessity to know the structure of the underlying dataset in order to create the correct query pattern.
One of the two directions of the research presented in this paper is to enable non-expert users to interrogate RDF databases through the easier Keyword Search paradigm, which expects the use of a bag of words as a representation of the information need.
Among the di culties regarding keyword search systems, as described in
[
Data Citation Today data has become fundamental for building and
interpreting new scienti c results. Citations are one of the most signi cant tools used
in the creation and propagation of knowledge and one of the basic means on
which scholarship and scienti c publishing rely. Citations permit to identify the
cited material; to retrieve it; give credit to its creator; date it; assign
responsibility or ownership [
Outline Section 2 reports the related works in the eld of keyword search and data citation in the eld of RDF graphs. In Section 3 we describe the general architecture of the two modules composing the keyword search and citation system we are developing. Finally, Section 4 describes some future problems and directions that will be tackled in the development of the system. 2
Keyword Search has been extensively studied in the context of structured
databases such as Relational DB and Knowledge Bases. Good reviews about these
topics are [
Regarding data citation, [
Regarding RDF datasets in particular, among the di erent proposals we nd
a nano-publication model where a single RDF triple is made citable via
annotations [
As of today, there is not a unique system that can enable users to easily extrapolate data from RDF datasets and to generate the citation for these data. 3
In general, as depicted in Figure 2, keyword search systems present two modules: an o -line module, performing operations on the dataset before the user's query arrives, and the on-line module.
In our implementation, the o ine module extracts subgraphs from the RDF
datasets with an algorithm called TSA [
Once the user's keyword query arrives, we can leverage on the index to
perform a rst fast ranking of the graphs using BM25 and then taking the top k
element of the ranking (e.g k = 1000). In this way, we obtain in the rst
rankings the subgraphs that are more relevant to the information need of the users.
Taking only the top k graphs of the ranking limits the space of potential answers
and limits the required computation time. Subsequent operations on the top-k
graphs include merging of graphs (when signi cant overlappings are present),
pruning (when the presence of triples without keywords in the outer part of the
graphs are detected) and a re-ranking using an MRF function [
Thanks to this division of operations in an on-line and o -line phase we performed queries over databases of ten and even hundreds of thousands of triples (LinkedMDB, LUBM, BSBM, IMDB, Dbpedia). 3.2
Once that the user has obtained the results for her keyword query, she may want to choose one or more of the answers that she perceives closer to her information need, and create a citation for that piece of information.
The role of the citation system is to automatically create the data citation associated with the graph chosen by the user without her direct intervention in the process. Since RDF datasets lack the structure of RDB and XML databases, it is necessary to rede ne some concepts and problems of data citation, as for example the identi cation of the citable unit (the minimal element inside the graph that can be cited), the granularity of the citation (the kind of structures inside the graph that can be identi ed as citable).
Considering the nature of the system we are developing, the citation system
should be able to cite subgraphs of di erent dimensions, composed of one or more
triples. One possible approach can be derived from the work proposed in [
The general structure of a system using nano-publication is depicted in gure
3. Ideally, in the system, every triple in the graph presents an associated
nanopublication. The di erent nano-publications can be combined together to create
the nal human-readable and machine-readable citation of the answer graph.
However, it is still to be studied how to aggregate the annotations over more than
one triple, or how to create a citation over a set of triples only partially provided
with nano-publications. An approach similar to the one presented in [
In this paper, we presented the two modules of a keyword search and citation system for RDF datasets that will enable users to automatically nd and cite data inside an RDF dataset. Our future endeavors will be directed toward the completion and betterment of the two models.
Regarding the keyword search system, we plan to work on the creation of virtual documents from the subgraphs, introducing elds to better exploit the information included in the IRIs. Moreover, we will explore the possibility to introduce query expansion to help the ranking performed by BM25 and the MRF-based function. We will also work on the provided answers that by now are subgraphs. We will perform entity extraction and the extrapolation of NL descriptions of the answers in order to help non-expert users to read them. We will also face the critical aspect of scalability. While our system scales well in time on databases of tens and hundreds of millions of triples, we saw that the performances of e ectiveness tend to decrease dramatically. We will study new methods to face the problem of big real-world databases.
Regarding the citation system, we will study methods to deal with the di erent granularity of the potential graphs to be cited. Among the major problems of the automatic citation of graphs, there is the semi-structured nature of RDF, which lacks the hierarchical nature of XML databases. New scalable techniques need to be designed in order to build human-readable and machine-readable citations, including the use of nano-publications and named-graphs.
This work is partially supported by the Computational Data Citation (CDCSTARS) project of the University of Padua and by the ExaMode project, as part of the European Union Horizon 2020 program under Grant Agreement no. 825292.