The Linguistic Linked Open Data movement aims to model linguistic data according to the Semantic licensing, and accessibility, to identify potentialities and limitations that limit its utility and exploitation.
CEUR ceur-ws.org
In the rapidly evolving landscape of linguistic research, the advent of Linguistic Linked Open
Data (LLOD) [
However, this endeavor is not without its challenges. The LLOD ecosystem is characterized
by a rich tapestry of resources, ranging from traditional linguistic databases to encyclopedic
knowledge bases like DBpedia and Wikidata, leading to heterogeneity in both content and
structure [
This article reports the quality assessment of the LLOD Cloud in terms of accessibility, use of open licenses, and amount of data. The LLOD Cloud1 counts more than 200 KGs in June 2024. From a preliminary evaluation of linguistic datasets modeled according to the LOD principles and attached to a scientific contribution indexed by Scopus, it resulted that several resources are still missing to the Cloud. In the direction of confirming it, the LLOD Cloud diagram is explicitly declared to be an ongoing project inspired by the LOD cloud diagram authored by Richard Cyganiak and Anja Jentzsch and includes open, available, and interlinked linguistic resources. As a consequence, it is expected that the diagram will incorporate an increasing number of resources over time.
At the current stage, the LLOD Cloud is organized in categories, which are corpora, lexicons&dictionary, terminologies, thesauri & Knowledge Base (KB), linguistic data categories, linguistic resource metadata, typological database (DB) and other. Categories are not balanced, as can be observed in Column # in Table 1. Moreover, we can observe the presence of a consistent portion of the Cloud categorized as Other. It might raise the question of whether the categorization is adequate and detailed enough. Further studies are required to verify if the current categories are aligned with linguistics’ expectations.
Methodology. Data to perform the quality assessment are retrieved by publicly accessible pages attached to resources published within the LLOD Cloud. We have downloaded all the resources in a single JSON file, and we retrieved the title of each KG, keywords to distinguish the LLOD category used to classify linguistic resources, sparql which report the SPARQL endpoint, if any, full_download or other_download to retrieve any download format attached to each resource, triples which model the amount of data, and the license. Besides returning the link of the SPARQL endpoint and the download format(s), the LLOD cloud also returns the status of each link. As a result, we can identify all the resources attached to a working link.
Starting from the quality dimensions defined by Zaveri et al. [
Results. The Python script to compute the quality dimensions along with the LLOD.json file are openly and publicly available online5. Quality assessment results are reported in Table 1 1LLOD Cloud: https://linguistic-lod.org 2Apache License: https://www.apache.org/licenses/LICENSE-2.0 3MIT License: https://opensource.org/license/MIT 4Creative Commons Licenses: https://creativecommons.org/licenses 5GitHub repository: https://github.com/isislab-unisa/LLODCloudQuality Persistent DOI on Zenodo: https://doi.org/10.5281/zenodo.13449868 clustered per LLOD categories. Per each category, we report: • column # - the number of LLOD resources in that category; • column OL - the number of LLOD resources attached to an Open License (OL), • column A - the number of LLOD resources attached to a working download mechanism, including to a working SPARQL endpoint; • column SE - the number of LLOD resources provided with a working SPARQL endpoint, and naturally a subset of the accessible resources; • column OL&A - the number of openly accessible LLOD resources, meaning that they have a working download mechanism and are attached to an OL; • column OL&SE - the number of LLOD resources openly accessible via a working SPARQL endpoint; • data columns reporting the total amount of data (column Tot.), the one accessible via a working download mechanism (column A), via a working SPARQL endpoint (column SE), via a working download mechanism and provided with an open license (column OL&A), and via a working SPARQL endpoint and provided with an open license (column OL&SE) Discussion & Conclusive Thoughts. LLOD resources are not uniformly distributed over the categories, as corpora, lexicons&dictionary, and other cover the majority of the LLOD. Less than half of the LLOD resources are attached to an OL, and in some cases, it is even deprecated, as happens by using CC BY-NC 2.0 license6 instead of the updated 4.0 version. This overall picture 6CC BY-NC 2.0: https://creativecommons.org/licenses/by-nc/2.0 is almost coherent with each category, with few exceptions, as the corpora category where 60% of resources are attached to an OL, while resources belonging to the linguistic resource metadata and typological DB categories completely miss OL. 40% of the resources are accessible via a working download mechanism, but only 5% of them are attached to a working SPARQL endpoint. The situation is even worse when we focus on openly accessible LLOD resources, as they drop to 14% while considering any download mechanism, to only four resources if we need an openly accessible LLOD attached to a working SPARQL endpoint. This results in an extraordinary amount of data being left untapped. While naming the worst case, LLOD in the lexicons&dictionary category sums up to 1G of triples, while 322MM can be downloaded, only 4MM can be openly and freely reusable, while no data can be accessed via SPARQL endpoints. In summary, there is a huge potentiality in the LLOD regarding the variety of resources and amount of data, but accessibility and the rare use of open licenses hinder their exploitation.