Matcha is an ontology matching system under development, designed to tackle long-standing challenges such as complex and holistic ontology matching. It incorporates all of the key algorithms from AgreementMakerLight (AML) over a novel broader core architecture, and includes several new algorithms. Matcha-DL augments Matcha to semi-supervised tasks, it uses trainable model to select and rank candidates proposed by Matcha. Matcha performed well overall, achieving the highest F-measure in 6 of the 18 distinct OAEI tasks and ranking in the top three in 9 others. Matcha-DL achieved the highest F-measure in 4 of the 5 semi-supervised BioML tasks.
A matching algorithm that uses Large Language Models (LLM) has been added to Matcha.
The strategy relies on the conversion of the entities’ labels and synonyms into embeddings,
followed by a computation of the cosine similarity between embeddings. The embeddings can be
obtained from any LLM model, although for this OAEI edition, we used the sentence-BERT [
Matcha’s novel translation module is a neural-based many-to-many multilingual translation method that uses an Encoder-Decoder Long Short-Term Memory architecture that consists of two complex recurrent neural networks that act as an encoder and decoder pair. This mechanism solves a sequence-to-sequence prediction problem between two given source and target ontologies’ languages.
The encoder maps the source language’s label to a vector representation that serves as input to the decoder, that then maps the vector back to a translation of the label in the target language. The label translations are then added to the lexicons of the original ontologies.
For Matcha-DL, a specific pipeline was developed that incorporates nearly all of Matcha’s matching algorithms to create an input for a dense neural network. Matcha-DL augments Matcha by learning to rank candidates produced by Matcha based on this input scores.
Matcha’s matching algorithms are described in Table 1.
The MELT [
Matcha-DL was mainly designed as a ranking model for local matching. It was however adapted to for global matching tasks by considering candidates produced by the Matcha core algorithm.
As Matcha is still under development, it is not publicly available. A public release is planned once the core development is completed, which is expected to be soon.
Matcha’s results for OAEI 2023 are summarized in Table 2, with the exception of the results for the BioML track, which are presented in Table 3, for both Matcha and Matcha-DL.
Matcha had good general performance, achieving the highest F-measure out of all systems in 8 of the 18 distinct OAEI tasks, while ranking in the top 3 in 9 others.
The participation in the complex track was hindered by the change in the definition of instances. In this year’s datasets the entities shared a local name with diferent prefixes, but Matcha’s algorithms rely on the entities being semantically equivalent – either by having the same URI or by being declared as owl:sameAs. 1https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2
Instance Matching
In regards to the Multifarm track, we had a character encoding issue that we were unable to resolve in time. In the Knowledge Graph track, Matcha only found matches for instances which was an unexpected result and requires further investigation.
In the Material Sciences and Engineering, some of the matches were between classes and object properties, which was puzzling considering that Matcha separates entities by type for the matching tasks, meaning there should be no mappings between diferent entity types. We suspect it could possibly be an issue with the ontologies’ encoding.
Matcha-DL, through the incorporation of a relatively simple training procedure to augment Matcha, demonstrates surprising results on the semi-supervised Bio-ML tasks. Remarkably, it achieves the highest F-score in four out of five tasks. Comparative analysis with the Matcha algorithm underscores the discernible advantage of incorporating training within the context of ontology alignment.
There is a general improvement in Matcha’s results when compared to the earlier version presented in OAEI 2022, although some further refinements are still required. Matcha achieved the highest F-measure in 6 of the 18 distinct OAEI tasks and ranked in the top three in 9 others.
This work was supported by FCT through the LASIGE Research Unit (UIDB/00408 /2020 and UIDP/00408/2020). It was also partially supported by the KATY project which has received funding from the European Union’s Horizon 2020 research and innovation program under grant agreement No 101017453. This work was supported in part by projet 41, HfPT: Health from Portugal, funded by the Portuguese Plano de Recuperação e Resiliência. Marta Silva was partially funded by FCT through the fellowship 2022.11895.BD.
FISH-ZOOPLANKTON
SNOMED-NCIT (Neoplas) – – – – – – – –