In the realm of data-driven businesses, structured data, being highly organized and easily understood by machines, is a valuable resource. IATE, with almost one million concepts storing multilingual terms and metadata, holds a large part of the textual knowledge of the EU. However, it can only be accessed lexically, and the database concepts stand alone. If IATE were taxonomized, i.e. related concepts linked up into knowledge graphs yielding a full-fledged ontology, its data could not only be consumed by linguists but would also become accessible by the machine readable SPARQL endpoint, which makes it a powerful resource for AI projects, particularly within SMEs that rarely have the means to create multilingual formalized knowledge.
Coreon team elevated a sub-domain of IATE terminology into a multilingual knowledge graph. We taxonomized a flat list of 425 concepts within the COVID sub-domain, benchmarking two approaches to tackle this task: automatically through a custom-enhanced off-the-shelf language model and a manual creation of the knowledge graph by a linguist expert. The automatically created knowledge graph was later revised by a human, corrections and time effort measured and compared with performance metrics of the manual approach. In this talk, we will dwell on performance and resource-saving advantages of our custom method and show how the achieved productivity rate can make the taxonomization of even large terminology databases economically viable.
We demonstrate empirically the effectiveness of our collaborative-robot approach in a typical industry use case scenario: using the resulting IATE/Covid graph for initialization of a Convolutional Neural Network (CNN) in a multilingual document classification task, we get a classification granularity that is not reachable by state-of-the-art models, such as non-initialised CNNs and zero-shot classifiers.
Keywords: auto-taxonomization, data modelling, knowledge graphs, SPARQL, data quality, machine learning, data visualization