Ontologies are an appropriate means to represent knowledge on the Web. Research on ontology engineering reached practices for an integrative lifecycle support. However, a broader success of ontologies in Web-based information systems remains unreached while the more lightweight semantic approaches are rather successful. The linked data initiative for example became a huge success during the last few years. Relying on the technologies of RDF and on ontologies as the appropriate means for the underlying vocabularies, linked datasets are one of the biggest and most actively used application areas of ontologies on the Web. We assume, paired with the emerging trend of services and microservices on the Web, new dynamic scenarios gain momentum in which a shared knowledge base is made available to several dynamically changing services and applications with disparate requirements. Our work is a step towards such a dynamic scenario in which an ontology adapts to the requirements of the accessing services and applications as well as the user's needs in an agile way based on ontology usage. Thus, our approach reduces the experts' involvement in ontology maintenance processes.
Ontologies are an appropriate means to represent knowledge on the Web.
Research on ontology engineering methodologies has come from describing the
scratch development of ontologies and reached practices for an integrative
lifecycle support. The ontology engineering discipline has changed from an individual
art towards a collaborative and distributed process with disparate skilled users
develop consensual models and distributed networks of ontologies[
As a next logical step one should await that, against the trend of the data
silos, the user holds and controls her data on her own. Paired with the emerging
trend of services and micro-services on the Web [
The general and personal motivation for this work consists of three core parts.
The first part is based upon our studies of the existing ontology engineering
methodologies. It represents the fundamental direction of this work. As a second
part, we derive from personal interviews with small and mid-sized enterprise
(SME) partners of the project Corporate Semantic Web, that they look for
a lightweight and dynamic process for ontology maintenance which minimizes
the need for ontology experts to be present. We explicitly focus this scenario,
however, we respect that there are enterprise settings as well which need and deal
with heavyweight ontology engineering processes. On the whole, our idea meets
the gap, which we identified as the result of the study of ontology engineering
approaches and which has been also identified by others, such as [
Agile ontology maintenance or in other words dynamic ontology evolution,
is an open problem in the research community. It is embedded in the hot
research field of ontology dynamics in general, which gained momentum since the
classical ontology engineering methodologies reached a mature state. Dynamic
ontology evolution is currently addressed under different scopes – the
domainoriented scope and the application- or usage-oriented scope. The most important
representative for research on the first scope is the work of Zablith[
To clarify this motivation we will briefly come up with a simple running example for the problem which we want to solve. Consider a company’s knowledge base which includes information about the employees. In the beginning only personal information have been collected conforming the friend of a friend (FOAF) vocabulary. One service uses the knowledge base which generates lists of employees with certain interests. Each time a new service is bound to the knowledge base, such as a service for displaying absent employees or information about the income (e.g. for the accounting), the maintainer of the knowledge base has to find out which facts, in the sense of the T-box of the ontology, are missing and how she can easily adopt the current T-box and possibly the A-box as well to these new application requirements. It is also possible that separate services require the same information represented in different vocabularies (e.g. foaf:name vs. myvocabulary:name) which yields the conflict for the maintainer whether to replace the present representation or to model a mapping between both. The decision for either the first or the latter depends on several criteria, such as the computability of the ontology for complex reasoning or obsolete and unused information. 1.2
Regarding the application of linked data or rather applications that use linked
datasets, an interesting research gap appears. The information which were
collected by classical Web usage mining techniques to improve Web pages in a
user-oriented way do not work out. The well-known principles of sessions, paths
or click-through do not exist in the same fashion for Web data as they exist for
Web pages. However, the improvement of the quality of a dataset with reference
to the user’s needs has to be in focus of each single dataset host. This problem
can be observed very well in the case of the DBpedia[
By now the DBpedia maintainers started to apply features which allow the user community to perform changes on the ontology and the mechanisms for instance population directly. However, it is an interesting question as well how our approach could help the user community in maintaining the DBpedia ontology in the future.
This motivating examples and the general problem description yield the central research questions of our work: 1. How does a methodology for ontology maintenance in an agile environment look like? We search for a process which puts less emphasize on the initial development of an ontology but more on the ontology usage and evolution. 2. Can we reduce the necessary influence of human experts in the ontology maintenance process by tracking feedback about ontology usage? In this case our work searches for a formal model that allows the analysis of ontology usage for ontology evolution purposes. 2
Our work is following the principles of design science research. Initially we started by a comprehensive analysis of the state of the art in ontology engineering. To our best knowledge we achieved an integrative overview of the different research directions and methodologies in this field. Thus, we identified one open problem to solve – the problem of agility and dynamics in the ontology engineering process. After this our work concentrated on defining the related context of our proposed solution explicitly. That includes a characterization of the type of ontologies we are addressing and an exemplary application scenario. The single parts of our solution were and are designed and developed stepwise in the following, before we finally will evaluate the theoretical saturation of our technologies in a multi-perspective way so it is possible to border its applicability from other approaches in a pragmatic way. 2.1
Altogether, this work aims at a multi-layered contribution as it is depicted in Figure 1. From the underlying theories we derive a methodology for agile ontology life cycles. Then we design and implement the necessary methods and tools for the core phases of the methodology and finally evaluate both layers, the methodology as well as the methods and tools, multi-perspectively.
The thesis will provide a proper understanding of the problems of agile ontology maintenance and propose the methodology which helps tackling these problems. Since the methodology differs from state of the art approaches in ontology engineering it proposes several innovative process steps which will be supported by the above mentioned methods and tools. Another and important contribution of our work is that the Semantic Web usage mining approach closes the gap between classical Web usage mining techniques and the new technical and organizational issues when Web data instead of Web pages is regarded. The work will not only elaborate on these issues but it will also present a practical solution to tackle them.
Application of ONTOCOM to
COLM
Comparative Discussion of
COLM Semantic Web Usage Mining
Real World Analysises
DBPedia Case
Study Statistical Heuristic Method for
Ontology Change
Recommendation
Corporate Ontology Lifecycle Methodology ʹ COLM tsen sno Ontologies & rem and itadn Ontology Engineering ieuqR Fuo
Web Usage Mining
Statistical Heuristics Three parts are the building blocks of our proposed solution for the problem stated so far: (1) A methodology for agile ontology engineering, (2) a method for Web usage mining in the context of the Semantic Web, and (3) a method for using statistical heuristics for ontology change recommendation. The latter part is in a preliminary design state, while the methodology and the usage mining method have already been developed and matured. The Corporate Ontology Lifecycle Methodology (COLM) reflects the agility of knowledge engineering processes and brings in application dependency. We define it as an agile ontology maintenance methodology since it is focused on continuously evolving ontologies in an application-dependent context. To clarify which process steps are more expert-oriented and thus need higher human involvement and those which need less, COLM consists of two different cycles, namely the engineering cycle (high involvement) and the usage cycle (less involvement). The overall goal is to use an intuitive reporting of tracked usage information which indicates the necessity of change.
As depicted in Figure 2, the process starts at the selection / development / integration phase. The result of this phase is an ontology, which is validated within an application-dependent context. If it is approved that the ontology suites the requirements it is deployed to be in use and it is populated. Throughout the whole feedback tracking phase, formal statements about users’ feedback and behavior are recorded and finally a reporting of this information is performed. The usage cycle is left if any necessary change has been detected and the knowledge engineers evaluate the weaknesses of the current ontology. 3.2
In this section we describe our approach to analyze server log files of (linked)
data endpoints with the goal to retrieve information about the usage of the
dataset and its underlying ontology. The log files contain information in the
extended common log format[
Listing 1.1. Examplary log entries of the DBpedia dataset
Even though our log file analysis collects general statistical information about
the number of requests, the number of different host, the peak access time slots,
and the user agents amongst others, it is primarily intended to provide
information about the returned results of the endpoints with reference to the users
requests. It is easy to see that the primitive observation of HTTP error codes
and response sizes does not yield the appropriate information about the size and
quality of the result sets or any conclusions about ontology usage on a concept
level. This aspect is why our approach differs from the work done in this field
by M¨oller et al.[
The analysis is performed on demand and not at the real runtime of the requests. That means we re-run each single call from the log files against a mirror of the DBpedia dataset, which shares the appropriate dataset version with reference to the date of the currently analyzed log file. SPARQL queries are performed against the mirror server as they have been performed against the real endpoint. The requests for single resources are reorganized as SPARQL queries following the simple schema noted in Listing 1.2. 1 SELECT * WHERE { <http :// dbpedia . org / resource /[ resourceidentifier ]> ? property ? hasValue } Listing 1.2. Generated SPARQL query related to a simple resource request from the log file
We partition each SPARQL query into its atomic parts – namely (1) patterns, (2) filters and (3) triples – and check each of these parts individually by sepcificly generated SPARQL queries. So finally we collect the following information about the requests: (1) Which queries are executed? (2) Which queries contain errors? (3) Which queries return a non-empty result set? (4) Which query patterns exist? (5) Which query patterns return a non-empty result set? (6) Which filters are used? (7) How do filters effect on the size of result sets? (8) Which triples are requested in queries? (9) Which triples do return a non-empty result set? (10) Which entities appear as subject, predicate or object? 3.3
Heuristics are guesses about efficient rules that can be used for problem solving. They rely on practical experience in certain types of problem-solving activities. In our case the problem is to evaluate the report which is generated by the usage mining method and recommending changes to be performed to the ontology and the dataset. A central question this heuristic approach should answer is whether and when the usage of an ontology primitive as a subject, predicate or object for example is significant. In this special case that has to respect the factor that the total number of properties which may be used as predicates is much smaller than the total number of classes and instances which may be used as subjects and objects in queries.
We are currently working on a combination of an algorithmic solution that relies on the retrieved statistics and, in the optimal case, on ontology engineering best practices and patterns, to recommend necessary ontology changes. Ontology changes can be recommended on the structural level of the T-box or on the instance level of the A-box, the dataset. 4
Evaluating design science artifacts is a complex thing. That is why we decided to evaluate our approach multi-perspectively with focus on qualitative evaluation methodologies because these are less strict methods with emphasis on stressing new approaches and ideas influenced by subjectivity and diverse research. To some extend the evaluations will focus the individual parts of our work to prove their applicability and pragmatic validity. But, we will also perform an evaluation that puts the things together and provides an integrative view to our solution.
The applicability of the developed methodology COLM will be evaluated in a comparative discussion with other ontology engineering methodologies. The most feasible model seems to create a goal-free and a goal-based evaluation as the basis for this discussion. The central ambition of the goal-free evaluation methodology is to compare different solutions for the same use case against certain criteria. A goal-based evaluation gives a qualified view on the achievement of a single solution, in reference to the recommendations and requirements of the implemented artifact, raised by the developers. In the end this evaluation should proof whether we succeeded in developing a methodology for ontology maintenance in an agile environment (research question 1).
Since the ONTOCOM[
To prove the validity and applicability of our analysis method and the associated heuristics for change recommendation we will perform and document several exemplary executions of them on well-known and widely-used live datasets, such as the Semantic Web Dog Food Corpus or DBTune2.
As it was mentioned before, we also want to evaluate our approach integrative. This will be done by a case study. The case study evaluation is aimed at the appraisal of how a definite process can be enforced referring to a given process description. In this special case dos that mean, that we want to convince practitioners of the applicability of the COLM methodology, the Semantic Web usage 2 http://dbtune.org/ mining, and the statistical heuristic change recommendation method to the task of ontology maintenance. The case study is set up as a kind of laboratory experiment which means that we observe the actions by the DBpedia development team when they perform maintenance activities on the DBpedia ontology and the DBpedia dataset. 5
In this paper we reported on our ongoing work towards agile ontology maintenance. A presentation of the general approach and research methodology was followed by a detailed description of the concrete contributions which we will achieve when this work is finished. The paper closes with an overview of the aimed evaluations.
The fundamental work for the definition of the research problem and its related context has already been done as well as the development of the first two of three major contributions – the methodology for agile ontology engineering and the methods for Semantic Web usage mining. At the moment we are working on the heuristic model for ontology change recommendation. In parallel we are performing analysis by use of our method on a massive amount of data from the DBpedia dataset and the Semantic Web Dog Food Corpus3. That data forms the basis for some of our evaluations which will be completed afterwards.
During the work on this thesis we detected several other related interesting research problems. The Semantic Web usage mining approach for example, which we presented in this report shortly, has a broad range of possibilities to be extended by the application of graph analysis methods. That could provide an interesting insight on the usage of Web data in form of so called heat maps of the RDF graphs or other visualizations that base on our metrics and statistics. An intersting discussion could also be why the approach of query observation and analysis has never been done in the world of databases. To our best knowledge, we were not able to find a comparable work in that area.