Alin is a system for interactive ontology matching that has been participating in all OAEI editions since 2016. In this new version, we improved the non-interactive version of the tool. We run it like the interactive version but do not use an expert in the interaction; we use ChatGPT in its place. ontology matching, Wordnet, interactive ontology matching, ontology alignment, interactive ontology alignment, lexical analyzer, ChatGPT 1. Presentation of the system Due to the advances in Information and Communication Technologies (ICT) in general, a large amount of data repositories became available as valuable assets for enabling integrated data exchange platforms across organizations. However, those repositories are highly semantically heterogeneous, which hinders their integration. Ontology Matching has been successfully applied to solve this problem, by discovering mappings between two distinct ontologies which, in turn, conceptually define the data stored in each repository. The Ontology Matching process seeks to discover correspondences (mappings) between entities of diferent ontologies, and this may be performed manually, semi-automatically or automaticall1y].[ The interactive approach, which considers the knowledge of domain experts through their participation during the matching process, has stood out among semi-automatic ones [2]. A domain expert is an expensive, scarce, and time-consuming resource; when available, however, this resource has improved the achieved results. Nevertheless, there is still room for improvements2][, as evidenced by the most recent results from the evaluation of interactive tools in the OA1EI (Ontology Alignment Evaluation InitiativeA).lin[3][4][5][6] is a system for interactive ontology matching which has been participating in all OAEI editions since 2016.
CEUR
ISSN1613-0073 actions, the better. In the interactive process, a key step is selecting mappings for the expert. An improvement in the alignment process would occur if we could find a way to reduce the number of mappings in this set without lowering the F-Measure, or at least reduce the number of mappings at a greater rate than the decrease in the F-Measure.
Alin participates not only in the interactive track but also in the anatomy track and the conference track, both non-interactive. Until last year, this participation occurred as follows: before selecting mappings to be presented to the expert,Alin would select some that went directly into the final alignment. Participation in the non-interactive tracks consisted solely of these automatically selected mappings. This year, we changed this. In addition to the automatically selected mappingsA,lin submitted the mappings chosen to be shown to the expert to ChatGPT. Therefore, Alin’ final alignment is now composed of both the automatic mappings and those approved by ChatGPT.
During its matching process,Alin handles three sets of mappings: (i) Accepted, which is a set of mappings definitely to be retained in the alignment; (ii) Selected, which is a set of mappings where each is yet to be decided if it will be included in the alignment; and (iii) Suspended, which is a set of mappings that have been previously selected, but (temporarily or permanently) ifltered out of the selected mappings.
Given the previous definitions, Alin procedure follows 5 Steps, described as follows: 1. Select mappings: select the first mappings and automatically accepts some of them.
Detailed in the ’Specific techniques used’ subsection below; 2. Filter mappings: suspend some selected mappings, using lexical and semantic criteria for that. 3. Ask domain expert: accepts or rejects selected mappings, according to domain expert feedback; 4. Propagate: select new mappings, reject some selected mappings or unsuspend some suspended mappings (depending on newly accepted mappings); 5. Go to step 3 as long as there are undecided selected mappings.
All versions ofAlin (since its first OAEI participation) follow this general procedure.
In this year’s version, in item 3, for the non-interactive tracks, the participation of the domain
expert was replaced by consulting ChatGPT.
1.2. Specific techniques used
• Step 1. Alin employs a blocking strategy where it does not consider data and object
properties from the ontologies at this step. It selects only concept mappings based on
linguistic similarities between previously standardized concept nameAs.lin automatically
accepts mappings with standardized names that are synonyms, using WordNet and
domain-specific ontologies, such as the FMA Ontology in the Anatomy track, to identify
these synonyms.
• Step 2. Alin suspends some selected mappings that exhibit low lexical and semantic
similarity in their entity names, removing them from the set of selected mappings. We
use the Jaccard, Jaro-Winkler, and n-gram lexical metrics to calculate the lexical similarity
of the selected mappings. We also used a semantic metric called thAelin metric. These
suspended mappings can be further unsuspended later, returning to the set of selected
mappings, as proposed in [
• Step 5. Go to step 3 until there are no selected mappings. 1.2.1. Mapping anti-patterns An anti-pattern mapping can be a logical inconsistency, a construction constraint on the ontology, or an alignment constraint. An ontology may have construction constraints, such as a concept cannot be equivalent to its superconcept. The alignment between two ontologies can have constraints. For example, an entity of ontology cannot be equivalent to two entities of the ontolog y ′. Anti-pattern mapping is a combination of mappings that generates a problematic alignment, i.e., a logical inconsistency or a violated constraint.
Alin participates not only in the interactive track but also in the anatomy track and the conference track, both non-interactive. Until last year, this participation occurred as follows: before selecting mappings to be presented to the expert,Alin would select some that went directly into the final alignment. Participation in the non-interactive tracks consisted solely of these automatically selected mappings. This year, we changed this. In addition to the automatically selected mappingsA,lin submitted the mappings chosen to be shown to the expert to ChatGPT. Therefore, Alin’s final alignment is now composed of both the automatic mappings and those approved by ChatGPT.
Alin is available2 as a SEALS package (It can be run with MELT).
We will compare Alin’s participation this year and last year in the conference track and the anatomy track, both non-interactive. We will not show the results from the interactive track, as the change this year only produced results in the non-interactive tracks.
In Alin’s participation in the Anatomy track (Tables1 and 2), there was a substantial gain in recall and a much less significant loss in precision. As a result, there was a significant gain in
F-measure. Last year, Alin ranked fith in F-measure among eight tools, and this year it ranked third among twelve. But this result came at the cost of a much longer runtime.
In Alin’s participation in the Conference track (Tables3 and 4), there was also a significant gain in recall but a substantial loss in precision as well—though less than the gain in recall—which resulted in an increase in F1-measure. Last yearA,lin ranked fourth in F1-measure among nine tools, and this year it ranked first among the same nine.
When comparing the interactive execution ofAlin with its non-interactive execution (Tables 5 and 6), we see that the interactive execution yields better results than the non-interactive one. This occurs mainly when the expert has a zero error rate. When the error rate reaches 10% in the Anatomy track, the non-interactive execution performs better. In the Conference track, this happens when the error rate reaches 20%.
This new version of Alin uses ChatGPT to simulate the specialist’s participation in noninteractive tracks. Its use resulted in a gain in recall, but a loss in precision—smaller than the gain in recall—which led to an overall improvement in F-measure in both the Anatomy track and the Conference track. In both tracks,Alin improved its ranking in the F-measure.
During the preparation of this manuscript, the authors used Grok (xAI) and ChatGPT (OpenAI) to assist with text translation, grammar and spelling correction, and improvement of readability. The authors reviewed and edited all outputs and take full responsibility for the content of this publication. OM-2017: Proceedings of the Twelfth International Workshop on Ontology Matching, volume 2032, 2017, pp. 13–24. [6] J. Silva, K. Revoredo, F. A. Baião, J. Euzenat, Interactive Ontology Matching: Using Expert Feedback to Select Attribute Mappings, in: CEUR Workshop Proceedings, volume 2288, 2018, pp. 25–36. [7] A. Guedes, F. Baião, R. Shivaprabhu, Revoredo, On the Identification and Representation of Ontology Correspondence Antipatterns, in: Proc. 5th Int. Conf. Ontol. Semant. Web Patterns (WOP’14), CEUR Work. Proc., 2014. [8] A. Guedes, F. Baião, K. Revoredo, Digging Ontology Correspondence Antipatterns, in: Proceeding WOP’14 Proc. 5th Int. Conf. Ontol. Semant. Web Patterns, volume 1032, 2014, pp. 38––48. [9] Results for oaei 2024 - anatomy track, 2024. URL:https://oaei.ontologymatching.org/2024/ results/anatomy/index.htm,laccessed: 2025-11-14. [10] Results for oaei 2025 - anatomy track, 2025. URL:https://oaei.ontologymatching.org/2025/ results/anatomy/index.htm,laccessed: 2025-11-14. [11] Results for oaei 2024 - conference track, 2024. URL:https://oaei.ontologymatching.org/ 2024/results/conference/index.htm,laccessed: 2025-11-14. [12] Results for oaei 2025 - conference track, 2025. URL:https://oaei.ontologymatching.org/ 2025/results/conference/index.htm,laccessed: 2025-11-14. [13] Results for oaei 2025 - interactive track, 2025. URL:https://oaei.ontologymatching.org/ 2025/results/interactive/index.htm,laccessed: 2025-11-14.