4 Alin is a system for interactive ontology matching. The ALIN version participating in OAEI 2020 applies natural language processing techniques (NLP) to standardize the concept names of the ontologies that participate in the matching process. As Alin selects through semantic and lexical metrics many of the mappings that the domain expert evaluates, we hope that the standardization of the concept names will improve the selection of the mappings and thus the generated alignment. This article describes the participation of Alin at OAEI 2020 and discusses its results.
however, this strategy has achieved results that are superior to automatic
(noninteractive) strategies. Nevertheless, there is still room for improvements [
Interactive ontology matching systems select mappings for domain expert evaluates. Alin selects many of these mappings through semantic and lexical metrics. As the concept names of the ontologies are not standardized, these metrics may return lower values than would be the case if they were standardized. This smaller metric may cause Alin not to select these mappings for evaluation by the domain expert. In its 2020 version, Alin proposes Natural Language Processing (NLP) techniques such as the development of regular grammars (in reality its equivalent regular expressions) and context free grammars along with their respective lexical analyzers (scanners) and syntax analyzers (parsers), for the concept names of the ontologies to be matched. The use of these NLP resources (scanners and parsers) makes it possible to translate di erent patterns used in the two ontologies into a unique one. This standardization allows Alin to select better mappings for the domain expert to evaluate.
To do the standardization, Alin will have a new phase before the execution of the program. In this phase, an NLP expert develops, manually, grammars to the concept names of the ontologies and their respective scanners and parsers. Alin uses these scanners and parsers during the execution of the program. This new phase is possible in an interactive ontology matching system because: 1. We know before the program runs which ontologies it will match, as we need to look for experts in the domain of ontologies to interact with the program; 2. The process of searching, meeting, and scheduling a day available for the expert to participate in the process can take a long time, probably a few days.
We can use this time of a few days until the execution of the program to develop the necessary grammars, scanners, and parsers for the ontologies. In this version of Alin, the authors of this paper played the role of the NLP expert. 1.2
During its matching process, Alin handles three sets of mappings: (i) Accepted, which is a set of mappings de nitely 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 5 Available at http://oaei.ontologymatching.org/2020/results/interactive/index.html, last accessed on Oct, 23, 2020. been previously selected, but (temporarily or permanently) ltered out of the alignment.
Given the previous de nitions, Alin procedure follows 5 Steps, described as follows: 1. Select mappings: select the rst mappings and automatically accepts some of them. We explain the selection and acceptance process below; 2. Filter mappings: suspend some selected mappings, using lexical 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 back to 3 as long as there are undecided selected mappings
All versions of Alin (since its very rst OAEI participation) follow this
general procedure. In this 2020 version, Alin includes a new step where an
NLP expert develops grammars, and their respective scanners, and parsers to
the concept names of the ontologies. Alin uses these scanners and parsers to
standardize the concept names of the ontologies and thus improve the generated
alignment. The new step can lead to, for example, correcting spelling errors and
unifying di erent spellings for the same concept name. More detailed examples
of possible standardization of concept names are presented in [
Alin applies the following techniques:
{ Step 1. Alin runs the scanners and the parsers for each concept name of the
ontologies, modifying it and standardizing it. Alin uses a blocking strategy
where it discards all data properties and object properties of the
ontologies. So, in this step, Alin selects only concept mappings, using linguistic
similarities between the concept names. Alin automatically accepts concept
mappings whose names are synonyms. Alin uses the Wordnet and
domainspeci c ontologies (the FMA Ontology in the Anatomy track) to nd
synonyms between entities.
{ Step 2. Alin suspends the selected mappings whose entities have low lexical
similarity. We use the Jaccard, Jaro-Wrinkler, and n-gram lexical metrics
to calculate the lexical similarity of the selected mappings. We based the
process of choosing the similarity metrics used by ALIN on the result of
these metrics in assessments [
There are logical constraints which should apply to several ontologies. For example, an ontology may have construction constraints, such as a concept cannot be equivalent to its superconcept. An alignment may have other constraints like, for example, an entity of ontology O cannot be equivalent to two entities of the ontology O0. A mapping anti-pattern is a combination of mappings that generates a problematic alignment, i.e., a logical inconsistency or a violated constraint. 1.3
To this version, Alin used the scanners and the parsers we developed for the ontologies of the conference and anatomy tracks.
Alin is available 6 as a package to be run through the SEALS client. 2
Results Interactive ontology matching is the focus of the Alin system. If you compare the participation of Alin in 2020 and 2019 (Table 4), you will see an improvement in the quality of the generated alignment, showing the e ectiveness of the techniques used. 2.1
The use of NLP techniques led to an increase in the F-Measure of non-interactively generated alignments in the Anatomy track but stability on the Conference track (Table 1). 2.2
In the Anatomy track, Alin was better than LogMap in both quality (F-Measure) and total requests, but worse in both aspects than AML (Table 2). In the Conference track, Alin was rst in quality and third in total requests (Table 3). 6 https : ==drive:google:com=f ile=d=1ZM 3g0aOgU ha 9V ptU bqk9nmnkF Cl7L=view?usp = sharing Interactive Anatomy Track In this track, Alin had a decrease in the number of interactions with the domain expert and an increase in the quality of the generated alignment, showing that the use of the NLP techniques are e ective for this track (Table 4).
Interactive Conference Track In this track, Alin had an increase in the quality of the generated alignment but an increase in the number of domain expert interactions (Table 5). 2.3
The quality of the alignment generated by Alin depends on the correct feedback from the domain expert, as Alin uses this feedback to select new mappings. When Alin selects wrong mappings, the quality of the generated alignment tends to decrease. If we compare this year's quality decline with last year's, we see that this fall is more sharp (Table 6).
The run time of Alin this year was shorter than last year (Table 7). In an Intel I5 with 10Gb reserved to Alin, Alin has run 20% faster this year than last year. The execution in OAEI had a reduction in the run time, but other systems also had this reduction. So this di erence may be due both to modi cations made in Alin and to changes in the computational environment.
Year Precision Recall F-measure Total Requests
Year Precision Recall F-measure Total Requests 2016 2017 2018 2019 2020 Evaluating the OAEI 2020 results, Alin has improved the quality of the generated alignment in the interactive track. However, an increase in the user error rate led to a slight worse alignment. Finally, the number of interactions with the expert was relatively stable since last year, with a slight increase (from 228 to 233 requests) in the Conference track and a slight decrease (from 365 to 360 requests) in the Anatomy track.
Another consideration is that this version of Alin generates the need for a new expert involved in the process, to develop artifacts (scanner, parser) required for scanning and parsing the name of the concepts. This NLP expert may not always be available, but if he is, the results have shown that his work can improve the quality of the generated alignment. Alin 2020 used NLP techniques to improve the standardization of the concept names of the ontologies to be matched. They have been e ective in increasing the quality of the generated alignment while being relatively stable with regard to the number of requests to the user. Alin had a decrease in run time but a more sharp fall in the alignment quality when the domain expert makes mistakes. An assumption that Alin now assumes with the inclusion of NLP techniques is the need of a scanner and a parser for the ontologies involved in the matching.