AgreementMakerDeep (AMD) is a new exible and extensible ontology matching system with knowledge graph embedding techniques. AMD learns from classes and their relations between classes by constructing vector representations into the low dimensional embedding space with knowledge graph embedding methods. The results demonstrate that AMD achieves a competitive performance in a few OAEI tracks, but AMD has limitations for property and instance matching. 1 Presentation of the system AgreementMakerDeep (AMD) is a new ontology matching system inspired by AgreementMaker [2, 3], AgreementMakerLight (AML) [7] and BootEA [19]. This year is the rst time that AMD participates in OAEI. It is designed with the main goal of higher e ciency for ontology matching problems by applying knowledge graph embedding methods.
1.1
Ontology matching aims to establish semantic correspondences or relationships
between concepts or properties of di erent ontologies [
AMD mainly utilizes string-based techniques [
Commons License Attribution 4.0 International (CC BY 4.0). 1 This paper is dedicated to the memory of Isabel F. Cruz. The architecture of AMD is shown in gure 1, including ontology parsing, string and lexical matching, knowledge graph embedding, model learning and candidate selection.
Ontology parsing owlready2 [
String and lexical matching We apply several text per-processing
techniques like stop-words removal and tokenization on class labels and annotations.
AMD uses the Base Similarity Matcher (BSM) [
the structure information of ontologies by relations translated from one class
to another class using a modi ed TransR [
Problem Formulation Given two ontologies O and O', we construct knowledge graph X and Y, and de ne the correspondence between two concepts as following triplets Tc;c0 = < c; r; c0 >, where r is the relation between c and c'. The problem is to nd mapping set M = f(cx; cy) X Y jcx cyg. In this study, we focus on one-to-one alignment and the relation between concepts is equality.
Let ~v(cx)= fv1; v2; :::vmg and ~v(cy)= fv10; v20; :::vn0g be two d-dimensional vectors sets of size m and n, we compute their distance with simple cosine similarity by d(~v(cx); ~v(cy)) = 1-sim(~v(cx); ~v(cy)) as follows: sim(~v(cx); ~v(cy)) =
X arg max cos(~v(cx); ~v(cy)) i=1 j We de ne the probability of the aligned labels between concepts cx and cy by p(cyjcx) as follows: p(cyjcx) = sim(~v(cx); ~v(cy)) (1) where
is the sigmoid function.
Knowledge graph embedding In AMD, we apply a modi ed TransR method
which translates concepts and relations into concept space and relation-specify
concept spaces, since there are multiple relations in the ontologies e.g subclassof
and disjointwith. In the original TransR, the projected vectors are de ned as cr =
cMr; c0r = c0Mr, and the score function as fr(c; c0) = kcr + r c0rk22 [
L =
X
X (cx;r;c0x)2S1 (cy;r;c0y)2S2 max(0; (fr(cx; c0x) 1) (f (cy; c0y) + 2) (3) where 1; 2; > 0 and 2 > 1, S1 is the positive triples set and S2 is the negative triples set. We set di erent values to ensure absolutely low margin loss scores in the positive triples for reducing the drift of the embedding and also keep the function of the margin-based ranking loss.
During the process that computes vectors, we need to generate negative
triples. Following the work of Sun et al. [
Candidate selection We select candidates based on a threshold of the classes knowledge graph embedding vectors similarity, and then compare the similarity with baseline if the pairs are in baseline result sets. 2.1
In AMD, we use stochastic gradient descent as the optimizer and con gure hyperparameters as listed: dimensions are set to 200 for the vectors. The learning rate is among f0.01,0.02,0.001g, and mini-batch is f5,10g. 1 = f0.01,0.05,0.1g, 2 = f0.5,1.0g. The number of nearest neighbors for negative sampling is f5,10,20g.
From the local evaluation results on the Anatomy track, the best parameter set is as follows: the learning rate is 0.01, mini-batch is 10, 1 is 0.01, 2 is 0.5 and 10 nearest neighbors for the negative sampling. 2.2
Our framework uses Python with Tenser ow2 and RDFLib 3, and is packed for SEALS using MELT. We use the best parameter set in local alignments for the OAEI submission, see section 2.1. 2 https://www.tensor ow.org/ 3 https://github.com/RDFLib
3.1
The Anatomy track results of AMD are shown in Table 1. AMD returns 1167 correspondences in 3 seconds. The result shows that AMD can be competitive among the top promising matching systems, especially in terms of runtime and precision. AMD is the second fastest system in this track and a slightly higher(0.004) precision than AML. The Conference track results of AMD are shown in Table 2. As expected, the performance of AMD in the conference track is not good, with the F-measure only slightly higher when comparing baseline method(StringEquiv). AMD shows a lack of ability to extract and match the properties in M2 and M3 evaluation variants. However, AMD has higher values in term of Precision in most tasks. AMD is able to complete two of the ve tasks with a runtime of 37 minutes, and AMD only returns class correspondences with a precision of 1.0.
Comments on the result 2021 is the rst time that AMD participates in OAEI, and performs promising results. Overall, the results show that AMD is able to complete several tasks in di erent domains on class-level matching in a timely manner. It is a fast system in most of tracks. Hence, we have shown that knowledge graph embedding is helpful to decrease computation time and that it leads to a competitive performance in term of F-measure in Anatomy and LargeBio tracks. AMD has consistently had higher precision than AML in a few tracks. However, AMD is still under development that it is only able to return class correspondences. Moreover, AMD has memory issues for large scale datasets and is not able to match properties and instances in the current stage. 4.2
The current development of AMD touches on several aspects. Besides considering
properties and instances matching, we will utilize joint embedding to combine
contextualized knowledge graph embeddings like coKE and BERT and additional
knowledge resources such as WebIsA [
In this paper, we have introduced a new ontology matching system called AMD. We adapted a modi ed transR model to t the ontology matching problem: thus, we learn low-dimensional embeddings for each class and relation to capture the hidden semantics of ontologies, rather than measuring the similarities between classes directly, as in other traditional systems. AMD makes full use of the textual and structure knowledge of ontologies. The results demonstrate the high e ciency and the promising performance of our proposed matching method as compared to other systems results in several tracks.