The FAIR Data Point has an increasingly important role in eforts to meet FAIR principles. It provides machine readable access to the metadata of diferent types of digital objects. In this paper, we focus on metadata of datasets. Since its first reference implementation, more tailored implementations have been developed and deployed in the Health Care and Life Sciences domain. However, a problem coming with these increasing amount of FAIR Data Point instances and the datasets published is the Findability of relevant datasets from the large volume of resources. For eficient finding of relevant datasets we need to exploit the richness of their metadata and a good ranking algorithm. In this paper we report the enhancements of the search and ranking capabilities of FAIR Data Point's reference implementation. Specifically, we improved its semantic search capability via creating association between class terms and the words frequently occur in the class description and labels. We also implemented a TF-IDF based ranking algorithm on the search results to present users the most relevant results. With these two enhancements, the FAIR Data Point can respond to a user's search request with higher coverage and present the list with the more relevant results based on the Term Frequency - Inverse Document Frequency (TF-IDF) metric.
FAIR Data Point (FDP) is a common approach to publish semantically-rich and
machineactionable metadata according to the FAIR principles[
More and more FAIR Data Points are being set up and running with various metadata of datasets from Health Care and Life Sciences domain are published to serve researchers. A detailed list of FAIR Data Point instances can be found in the FAIR Data Point HOME Server6, where, at the time of this paper’s writing, there were 41 active instances hosting metadatas of datasets and other types of digital objects.
An aspect that is currently under-investigated and important to increase the Findability of datasets published are improvements in the engines searching capabilities for relevant datasets. Dataset search is often complicated and ineficient when compared to a typical internet search, where algorithms use criteria of similarity between the potential keywords and the content and links included on websites. The current reference implementation of FAIR Data Point only allows for searching and ranking datasets in a primitive way. The main reason for this technological limitation is that links between datasets are still rare. This compromises the use of traditional web-based ranking algorithms.
In this work, we used the FAIR Data Point reference implementation as the basis for our
enhancement work. We implemented a semantic search framework for datasets that can extend
existing dataset search tools in two ways: 1. improving the semantic search capability over
metadata via association with frequent words occur in class labels and description in ontology.
2. ranking the search results by applying the Term Frequency - Inverse Document Frequency
(TF-IDF) [
To imkprove the search capability of the FAIR Data Point’s reference implementation, we designed a prototype including: • A Semantic Query Enhancer (SQE) component, which enhance queries by associating the user’s search keywords with terms occur in class labels and descriptions in the pre-loaded ontology. • A ranking algorithm based on TF-IDF metric to rank the results retrieved in the previous step.
This is done in the following steps as depicted in Figure 1
• index metadata from the ontologies
• retrieve and store associations for the search query words
• find documents for each associated words that has suficient relevance
• score and rank the documents, before returning them
2https://github.com/fair-data/fairdatapoint
3https://github.com/LOVDnl/fdp.lovd.nl
4https://repository.surfsara.nl/
5https://www.castoredc.com/
6https://home.fairdatapoint.org/
Ontology For ontology, we chose Thesaurus.owl and generate association by linking the
word to the class description. An example is given in the Figure 2 where “disease” is associated
with “pain”, because they occur in the same class description. The NCI Thesaurus [
On the portal of our ontology enhanced implementation7, a user can use the ontology enhanced search capability by clicking ”Switch to Ontology-based” link, as shown in Figure 3a. A simple comparison of the search capability between the reference implementation and our ontology enhanced implementation is given in Figure 3. Specifically, in Figure 3a, a search of the keyword ”disease” in the reference implementation gives 0 results. However in Figure 3b, the keyword ”disease” found 2 results. The reason of this diference attributes to the association between ”disease” and ”immunology”/”interleukin-1”, which co-occurred in the class description in the Thesaurus.owl ontology.
(a) Results of searching ”Disease” in the reference implementa
tion. (b) Results of searching ”Disease” in the Ontology Enhanced
implementation
Due to lack of user logs on the reference implementation of FAIR Data Point server, it is hard to apply a machine learning based ranking algorithm. These user logs supposed to contain the search keywords a user entered and the target results the user clicked. We plan to log the user behaviors at our enhanced FAIR Data Point portal to capture the keywords a user entered and the datasets the user clicked. With these logs, it is possible to train and apply a learning to rank algorithm. With more datasets being submitted and published to our running FAIR Data Point instance, a more detailed evaluation on the search performance in terms of precision and recall would be available.
In the existing setup, the singular ontology Thesaurus.owl is employed, given its status as a comprehensive ontology in the field of cancer research. However, it is worth noting that the incorporation of multiple ontologies is feasible, provided that technical challenges such as memory consumption are efectively addressed. In our upcoming implementation, a configuration option will be introduced to enable users to select ontologies based on their specific requirements.
This work was supported by SURF-DCC via the pilot:”Enhancing FAIR Data Point’s Search Capability as a FAIR Service v2.”