Task-oriented search scenarios go beyond retrieving information when a one-time perception of search tasks is neither possible nor sufficient. Such scenarios typically need further investigation, navigation or understanding of the search results. Formulating a search query is particularly difficult in case of distributed Linked Data sources, because they have many different relationships and vocabularies. Since users cannot realistically construct their intended query correctly at the first attempt, they need an environment in which they can iteratively refine what they are searching for. Therefore, this PhD thesis proposes an adaptive set of techniques and implements them for use cases in academics, industry and government to measure the effect on the user experience. We show that the set of techniques facilitates web applications in fulfilling task-oriented searches more effectively and that user interaction with search results indeed gradually refines search queries.
Typically, when users formulate search queries to find relevant content on the Web, they intend to address a single target source that needs to match their entire query. In cases when users want to discover and explore resources across multiple sources they need to repeat many sequences of search, check and rephrase until they have precisely refined their searches. The application of the Web of Data to search, makes it possible to extend basic keyword searches by describing the semantics of data and enables humans and machines to work together using controlled vocabularies. This enables distributing search tasks across datasets directly benefiting from a semantic description. Due to the high degree of mismatches between the structure of Linked Data and the variety in vocabularies across different sources, exploring distributed heterogeneous data sources is considered challenging.
Exploratory search covers a broader class of tasks than typical information retrieval
where new information is sought in a bounded conceptual area rather than having a
specific goal in mind. The users’ demand to discover data across a variety of sources
at once, requires searching facilities adaptive to their adjustments while they discover
the data that were just put at their disposal. In general exploratory search describes
either the problem context that motivates the search or the process by which the search
is conducted [
The general focus is the iterative exploration of linked data spread across different structural heterogeneous data sources. As there is no immediate suitable benchmark methodology for this model, it is necessary to rely on user-centered approaches and to develop reproducible automated machine approaches (using a gold standard). These approaches can be used to evaluate the application of the model in several prototypes which in turn allows us to observe how it enhances test-users search productivity and understanding of the data. 2
The generic methods and techniques developed in this PhD thesis find their application scenarios in various socio-economic relevant areas in academia, public sector and private sector. We give explain and motivate an example for each of the sectors. 2.1
Here the focus is bridging the walled garden of institutional repositories for ‘Science
2.0’. Much research data and publications are publicly available online, not only via
institutional repositories. The evolution of the Web to the Web 2.0 enabled a wide range
of lay users via wikis, blogs and other content publishing platforms to become the
main content providers. Combining information resources over the walls leads to a high
degree of mismatches between vocabulary and data structure of the different sources
[
This example integrates application data from many local governments in reusable
single purpose applications for ‘Smart Cities’. If local governments keep developing
(ad-hoc) data models and structures for this data over and over, it requires constant
revising the model of available data while in fact not being able to cope with newer
technologies and applications without heavily investing in new support infrastructure.
For example, instead of making a street event organization application only for a single
municipality, which outlines municipal services needed and permits required depending
on the type of event, governments develop an event organization application usable for
all municipalities in the region [
In the last example, the goal is to embed data visualizations in industry search
applications. In the industry, cases like those in the pharmacy-industry involve many partners in
the development of a product (e.g new medicine). Every partner focuses on providing
data for a different aspect such as the clinical trials, compounds and processes. It is
thus complex to build systems that integrate and align this variety of data. Typically
this data is very well structured or has high quality meta-data. Besides the
pharmacyindustry, also the media and entertainment industry can benefit from such a framework.
When recombining data from multimedia archives or social media for storytelling, new
hidden relations and trends among existing sources could be discovered by properly
describing and aligning them, enabling applications developers to design a whole range
of interesting and entertaining applications and visualizations [
Mostly direct querying approaches were tried and applications were often built around a limited set of supported SPARQL patterns. Furthermore, SPARQL queries are still hard for end users or even developers, despite GUIs and advanced query builders. Only in the last years vocabularies are getting streamlined and linked data is maturing. This leads to much more possibilities compared to traditional keyword search. Exploratory search in the front-end makes sense and transitioning from traditional web search and retrieval is changing. More and more web users and scenarios where exploratory search is beneficial appear (even though the paradigm is not new as such). The additional effort required for mapping, interlinking and maintaining data sources (i.e. as Linked Data), improves their re-usability and makes the methods and techniques for exploratory search immediately applicable. In the latter there are two scenarios: one where two data sources need be explored without interlinking them and the other where the effort is made: initial extra effort vs. reduced effort for implementing exploratory search. 3.1
We investigate how users find the information they need and gain insights about the data being under exploration through applications that enable them to interact with distributed heterogeneous data sources. The following questions is required to be addressed for attaining a set of techniques for exploratory search: – Can task execution be effectively facilitated by revealing relations between resources, i.e. adequately addressing the user’s intent? – To which degree does the additional interaction positively influence the relevance and precision of the search results? – How does a justification of the presented results influence the user’s certainty in getting closer to achieving the task’s goal? – How does the refinement of a search query gradually improve by interacting with its search results?
It is relevant to measure if and how well agreeing on semantics proves to be useful in tackling these issues. Our approach and evaluation illustrates how to apply semantic paradigms for search, exploration and querying. 3.2
Our research questions induce the following hypotheses: – Interacting with the search results refines and improves the result set because interaction with the result set makes the information contained in the initial search query more specific, leading to more and more targeted queries. – When exploring the data, indications such as facets, visualizations (charts, graphs etc.) reduce the number of steps to achieve a task’s goal. – Ordering of search results does not affect the search, neither in terms of steps needed, nor its precision.
Most of the works in literature about exploratory search, semantic search and distribution of queries across data sources deal with one or more aspects and are either focused on the front-end or the back-end. Typically they are limited to either a homogeneous dataset or they are purely focused on resolving the heterogeneity. In exploratory semantic search all these aspects need to be integrated. To the best of our knowledge there is no system that does all this. Nevertheless, one of the main contributions in this work is the distinct support for search scenarios where the revealed relation is one that the user was not aware of beforehand; besides describing methods and techniques for web developers and search applications on how to integrate exploratory search. However, there have been a lot of projects that cover multiple of these aspects playing an important role to make the whole work together. Therefore, we divide the related work section into two parts: (i) the front-end, search interfaces; and (ii) the backend, semantic search engines. The opportunities lie in adaptive techniques applicable to combinations of different linked data sources covering the entire work-flow from back-end to front-end without denormalizing the semantics along the way. 4.1
The set of tools focus on revealing relationships between resources and exploring them.
They contribute to distinct example solutions and implementations of adaptive and
intelligent web-based systems [
Recent developments demonstrate that Linked Data has arrived on the level of local
governments, public services and their target user group: citizens. Initiatives such as the
European Commission’s “Interoperability Standards Agency” (ISA) 1 enforce the use
of Linked Data and its data model RDF. Such data models are key for a formal semantic
representation of data resources. Semantic search is one of the main motivations behind
bootstrapping the Web into the Web of intelligent agents. Work on Semantic Web search
engines like Hermes [
Based on our experience in use cases in different domains (academia, industry and government) we identify and investigate a set of techniques for aligning and exploring data and verify that they are applicable in each of the domains. We generalize these techniques and iteratively refine them in an experimental setting where the data and queries are chosen carefully to highlight certain aspects (as depicted in the evaluation plan) to make the techniques applicable beyond the initial use cases we investigate. The goal is to optimize exploration techniques to the greatest extent. This involves detecting patterns in the data and defining a strategy for querying them accordingly, thereby balancing between common - and more rare queries fitting each scenario. 5.1
The techniques focus on generating views and abstractions, i.e. implement a query translation mechanism, accessible for end-users through services, and user interfaces. The other part focuses on aligning the data sources. Each of the use cases focuses on different aspect: The academic use case focuses on presenting the data to the users and turning them available for querying. The industry use cases implement translation techniques for the search tasks to queries. The government use case focuses on the semantic descriptions of the data to be able to query the data.
We developed a semantic model for searching resources in the Web of Data developed
for data related to scientific research (e.g. conferences, publications, researchers) [
Our approach leverages RDF, and the annotated semantic graph by relying on the fact that the vocabularies used in them can be linked. similar data of different source can thus be described in using the same terms, making it possible to explore these sources with the same queries. The user interaction with the RDF datasets occurs through a set of interfaces. Each interface facilitates the reuse, exposure and publication of digital research content as Linked Data. The interfaces bridge each of the components in the search infrastructure. 6
We elaborate on the evaluation methods and present intermediate results indicating the feasibility, effectiveness and usefulness of the techniques: – Case by case: the evaluation focuses on the use cases overall user perception and information retrieval quality (Eeffectiveness). Thereby we are testing both the (task-oriented) user experience and information retrieval aspects of each approach. We deduct as much as possible information out of these real-world proof of concept settings to address the research questions and hypotheses in. – Generic applicability: each hypotheses is evaluated directly and each of the research questions is address individually, in a perfect environment. Individual aspects are to be tested on a standardized collection and a standardized set of queries, changing only a single parameter to be able to test the hypotheses. Specifically we want to test the effects of returning results as a set rather than a list; test where two data-sources are being explored without interlinking them and the other where the effort is made; and the impact on the number of steps or time needed to complete a task when justifications are presented and cases when they aren’t.
In each of both cases, the approach is evaluated in two ways: (i) automated - by machines - after defining a suitable baseline for comparison (quantitative); (ii) user tasks - by observing user interactions with prototypes that implement the techniques (quantitative) and an accompanying user questionnaire (qualitative). Since the main purpose of the techniques is to facilitate users in exploring Linked Data on the Web, the evaluation of our approach is focused on both the end-users and the precision of the search results, as perceived by them.
Therefore, we investigate and define: – the characteristics, worth to be evaluated, of the data used in the experiments and – the baseline against which the search engine is evaluated.
Hereby the focus lies on information retrieval (IR) aspects which are important to quantify because it is inherent to any type of search (thus also exploratory search) and user-centered aspects. IR measures do not give the whole picture in exploratory search as they do in traditional query-centric search, in particular task-oriented, user centric, measures, are particularly useful evaluation criteria in exploratory search.
The processing of queries and mapping of keyword queries proved to be of promising
precision, given the complex and dynamic nature of the used datasets: a combination
of Linked Open Data and non Linked-data sources. We observed that searching by
keywords for resources increases the result set with more new relevant resources, while
it is on average as precise as expanding existing resources in the result set. The results
of a short survey[
The final interface, provided to the end-users, gave abundant and accurate
information about users, when the quality of the underlying alignment between datasets
has high accuracy and minimum sensitivity [
Considering that the implementation is still in the prototype phase, the potential
of a set of techniques to support visual and interactive search is well demonstrated and
2 http://www.informatik.uni-trier.de/˜ley/db/
3 http://wwww.colinda.org
understood by the target users. This relies mainly on the generic algorithm we developed
for revealing relations between Linked Data resources [
We aim to deliver the core building blocks for user oriented search engines and to facilitate exploring Linked Data, and ensuring their effectiveness by measuring: (i) the search precision; (ii) the support for re-usability of underlying data; and (iii) the degree of which they make search task execution more efficient. This PhD thesis investigates methods and techniques for web applications to support iterative refinement of queries for exploratory search with Linked Data. Overall, supporting such exploration on top of Linked Data: turns the potential of its exploitation more likely; and while allowing a larger group of users to discover Linked Data at the same time it increases the demand for this type of data, both in terms of context and semantics.
The enrichment of the main used data sources with Linked (Open) Data sources allows users to find a vast amount of resources implicitly related to them and thus initially not accessible. Facilitating exploration and search across semantically described distributed heterogeneous data sources is useful because it is still a laborious task for users to construct separate search queries for each of those services separately. We show how end-user applications facilitate accurately and iteratively exploring of linked data, without the need for a traditional ranked list of results. The set of techniques contributes to authenticity of the data it models and processes by guaranteeing that the final output towards the user has useful results in its domain of application. Because we stick with our approach close to the original structure of the data, this method is applicable to other domains if it is adequately structured by adapting the chosen vocabularies according to the datasets used.
The techniques contribute to users desiring to iteratively formulate precise searches and discovering new leads or validating existing finding across heterogeneous data without having to hassle with trial and error using traditional search engines. This will allow links to be revealed available but also to incorporate network structured data such as social and research data beyond the typical single user’s scope. This should lead to more fine-grained details facilitating users to obtain a more sophisticated selection and linking of contributed resources based on previous assessments and explored links.
I would like to thank my supervisors: Ruben Verborgh, Erik Mannens and Rik Van de Walle, for their support and the opportunity for the realization of this work. 4 http://www.resxplorer.org