Deliverable D5.5

[cbadenes]

Final Version of Ontology Learning and Matching Techniques with report

T5.2 Ontology learning

This task will exploit the information that is made available in research objects. Particularly important will be the information that can be extracted from research documents like published papers or technical reports and notes, and from the social network that is formed by researchers, their research objects and the content of these aggregations of scientific items. We will work on methods and algorithms to build up concepts and to identify the relationships between them, starting from this plethora of complex aggregated information, with the aim of building heavyweight ontologies, which may go well beyond the current lightweight ontologies that are being built in most ontology learning approaches. This will make use of the information extraction techniques from WP4:T4.4.
The general objective of this research will be to simplify the creation of domain ontologies from a set of seed research objects (a set of sample documents, scripts, social relationships between researchers, etc.) in the selected domain in this project. These domain ontologies should be adequately connected to an upper-level ontology network, which will be developed as well, reusing as much as possible existing ontologies in this area (e.g., BIBO [BIBO], EXPO [EXPO], SPAR [SPAR], etc.). Such an ontology network will be designed with the scope in mind of allowing the representation of scientific creativities, and a skeleton of it will be also built through the learning from a wide variety of research objects, including not only documents, but also datasets, workflows, scripts, etc. In any case, manual effort will be also required to complete and document this upper-level ontology network.
The learned ontologies will be used as a reference for the work of ontology-based information extraction in WP4:T4.4 and of the generation of ROSs in WP3:T3.1. The ROS will encode main technical initiatives of the input research document in terms of the technical elements (&concepts) and their relationships. For instance, one of the areas where we will pay special attention will be the one focused on methods and tools, which in many scientific areas is focused on describing the protocols used in a specific research work, and should aim at allowing reproducibility of the experiments. We will also focus on knowledge extraction from user-defined tags associated to research objects and their aggregated objects, following on current work in ontology learning from folksonomies.
In addition, extending existing work on social recommendation of research objects that UPM is undertaking in the context of the Wf4Ever project, we will be able to discover implicit relationships between different pieces of work that were originally not considered by the author in a basic literature exploration activity, what can increment creativity in the research.

T5.3 Interactive ontology matching for originality evaluation and creativity exploration

This task will work on techniques that explore similarities between research objects by means of exploring the similarities between the ontologies (including instances) generated as a result of the work described in T5.2, and by applying structured similarity evaluations between these aggregations of objects. In this context, scientific data should be better seen as claims rather than facts, and knowledge extracted automatically from documents and other artefacts should be seen as a skeleton set of information that summarises key ideas, and which allows researchers to explore the content of existing research objects in the process of their evaluation and scientific creativity.
An ontology mapping approach establishes which concepts in different ontologies correspond (equivalence)
to each other (and possibly other relations such as ‘more general’, ‘less general’, ‘disjoint’, etc). Notice
that, multiple terms are used to refer to similar technology: ontology alignment, ontology merging, ontology integration, ontology morphism, etc. The problem they all try to solve is interoperability between different knowledge bases (e.g. of different applications). The mapping makes it possible to make meaningful inferences about the union of the knowledge bases.
Many algorithms exist that create mappings between ontologies. Some approaches utilise linguistic similarities, while others use the semantics of the concepts (represented in the instances and relations of the concepts), or combine both approaches. We want to make use of existing techniques in the context of applying structured similarity evaluations between the aggregations of objects that are represented by Research Objects. In this context, scientific data should be better seen as claims rather than facts, and knowledge extracted automatically from documents and other artefacts should be seen as a skeleton set of information that summarises key
ideas, and which allows researchers to explore the content of existing Research Objects in the process of their evaluation and of the generation of scientific innovation.
Once the matching between Research Objects are obtained through the model comparisons, the degree
of overlap between different models can be used to determine other colleagues who are working on similar topics, to recommend colleagues that are working on similar topics to that they can both learn from their mutual creativity experience (then talking about collaborative creativity), or reflect about their findings.
We will also develop a model for representing hypotheses together with supporting evidence from research objects developed by a researcher, a group of researchers, or even by unknown or unconnected researchers and research groups. The model will be available in the form of a scientific workbench which enables scientists to explore the evidences, which have been obtained using ontology matching techniques, for which we will extend existing work in ontology matching in UPM. In order to foster novel evaluation and creativity generation, the workbench will record the subjective decisions of scientists about which hypothesis to accept, enabling them to share their assessments with other scientists working in the field. This should be valid across scientific domains. The ontology matching will be very useful in the similarity metrics WP3:T3.2, analogical reasoning processing in WP3: T3.4(i.e. retrieval, mapping and validation phases). This will also contribute to cross-source information extraction in WP4:T4.4, where similar objects from different sources will need to be identified.

T5.5 Evaluation methods and tools

An evaluation methodology will be put in place for the semantics tasks. We will integrate whenever possible available tools or implement our own software. The task will also run the different intrinsic evaluations which will help assess the progress of the technological development in the work package.