We present a demo of the GraFa faceted search and browsing interface over the Wikidata knowledge graph. We describe the key aspects of the interface, including the types of interactions that the system allows, the ranking schemes employed, and other features to aid usability. We also discuss future plans for improving the system. Online Demo: http://grafa.dcc.uchile.cl/
Faceted browsing [
To bridge this gap, we propose the Graph Facets (GraFa) system designed to offer faceted search and browsing over large, diverse RDF graphs. An important feature of the GraFa system is that it provides exact faceted views, meaning that the facets offered to restrict the current results set offer exact counts and are exactly those that will lead to non-empty results upon selection. However, result sets for very common types of (intermediate) queries – such as human or human AND gender male, etc. – can reach into the millions of entities. Computing the exact facets for such large results while maintaining interactive response times is technically challenging and does not appear to be well-supported by available faceted search tools. To improve scalability while maintaining efficiency, the GraFa system thus incorporates novel indexing schemes that, in an offline phase, pre-compute and store exact facets for large results sets.
In our paper accepted in the research track [
In the demo track, we propose to offer attendees of ISWC a live demo of
the GraFa system for performing faceted search and browsing over Wikidata.
We also wish to discuss possible features that could be added to the system in
future, other datasets or use-cases to which Grafa could be applied, as well
as to gain feedback and identify potential topics for collaboration. The demo
we plan to provide is publicly available here: http://grafa.dcc.uchile.cl/. In
this paper, we provide details on user interactions, details on the prototype, and
current limitations; for more information on performance, back-end, indexing,
usability, etc., we refer to our paper in the research track [
Figure 1 provides an overview of the user interactions that the GraFa system currently supports. The user is first presented with the option of searching by keyword (e.g., "nick drake") or by selecting a type IRI (e.g., wd:Q5 (human)). Each result set is associated with a list of facets from which the user may iteratively select further restrictions of the current results. We will now discuss each of these interactions in further detail.
For the initial keyword search, a ranked list of entities are returned that have matching keywords in their labels, aliases or de- Fig. 1: Overview of user interacscriptions; the properties corresponding to tions supported by GraFa such values must be configured in the system (in the case of Wikidata, we use rdfs:label, skos:altLabel and schema:description). Furthermore, a set of supported languages must be configured, where the labels, aliases and descriptions of these languages will be indexed, as available; the demo is configured for both Spanish and English.
Rather than search by keyword, the user can instead opt to perform a type search; here, types are defined as values for a configured list of type properties (in the case of Wikidata, we use wdt:P31 (instance of)). Given that users will often not know the required IRI of a type, the GraFa interface offers auto-completion on the labels and aliases of types in the graph, where suggested types are ranked according to a PageRank score and are annotated with the number of results with that type. For example, if a user types in the partial label hum, GraFa will suggest human (3595226 results) as the first result; if selected, GraFa will search for entities of type wd:Q5; on the other hand, if the user types in per, the first suggestion will be person (3595226 results), which if selected will also trigger a query for entities of type wd:Q5 (for which person is an alias).
Whether the user begins with a keyword or type search, GraFa will generate a list of resulting entities. For each entity in the results, its label, aliases, description and an image are displayed; the label is presented as a hyperlink that will dereference the entity IRI. In the case of keyword search, results are ranked by a combined query-relevance and PageRank score. In the case of type search, results are ranked purely by PageRank score. In either case, a list of facets are computed for the current results set. Each facet is a property and a list of values that at least one entity in the current result set is associated with. The facet view displays all possible properties, with a count of the number of entities with some value for that property. Upon selecting a property, the user can perform an auto-complete search on the label of a particular value, or can browse possible values in a drop-down list. Once a value is selected, a new result set is generated for the active conjunction of restrictions, where these results are ordered by PageRank; a new facet view is also generated. The user may continue iteratively adding facets until they are satisfied with the results, or until they reach a single result (facets with zero results are never offered). 3
We have implemented a prototype of the GraFa system for experimental purposes and to gather initial feedback and expressions of interest. The system uses Apache Lucene as a back-end store, managing (1) full-text search indexing for keyword and auto-complete prefix searches, (2) structured indexes for type and facet selection, (3) indexes for cached queries generating more than 50,000 results. The front-end is implemented as a Java servlet, with interactive autocomplete features being based on Javascript libraries. The source code is available from the following repository: https://github.com/joseignm/GraFa/.
We will demonstrate an instance of the GraFa prototype indexing a dump
of Wikidata. More specifically, the demo indexes the “truthy” dump of Wikidata
from 2017/09/13, containing 1.77 billion triples and 74.1 million entities. On a
machine with an Intel Xeon E5-2609 v3 CPU, 32 GB of RAM, and 2 2TB
Seagate 7200 RPM 32MB Cache SATA, indexing the data takes approximately
5 days, with the majority of time (4.5 days) taken to compute and index all
queries with more than 50,000 results [
In Figure 2, we provide an example screenshot of the GraFa demo where the user has searched for lighthouses on the Adriatic Sea, and is now considering further filtering the 14 available results by country.1 4
We see the current GraFa system as offering a baseline system for future
development, where indeed the current version has a number of limitations that we
have yet to address, including (1) support for datatypes and range queries, (2)
support for existential value queries, (3) support for class/property hierarchies
and potentially other forms of inference, (4) incremental updates. We note that
1 See http://grafa.dcc.uchile.cl/search?instance=Q39715&properties=P206%23%23Q13924
existing faceted browsing systems support some of these features, where, for
example, Broccoli [
Acknowledgements This work was supported by the Millennium Institute for Foundational Research on Data (IMFD) and by Fondecyt Grant No. 1181896.