Multi-view Embedding for Biomedical Ontology Matching?

Abstract

The goal of ontology matching (OM) is to identify mappings between entities from different yet overlapping ontologies so as to facilitate semantic integration, reuse and interoperability. Representation learning methods have been applied to OM tasks with the development of deep learning. However, there still exist two limitations. Firstly, these methods are of poor capability of encoding sparse entities in ontologies. Secondly, most methods focus on the terminological-based features to learn word vectors for discovering mappings, but they do not make full use of structural relations in ontologies. It may cause that these methods heavily rely on the performance of pre-training and are limited without dictionaries or sufficient textual corpora. To address these issues, we propose an alternative ontology matching framework called MultiOM, which models the matching process by embedding techniques from multiple views. We design different loss functions based on cross-entropy to learn the vector representations of concepts, and further propose a novel negative sampling skill tailored for the structural relations asserted in ontologies. The preliminary result on real-world biomedical ontologies indicates that MultiOM is competitive with several OAEI top-ranked systems in terms of F1-measure.

Keywords

• Ontology Matching
• Embedding
• Cross-Entropy
• Negative Sampling

-

In the Semantic Web, ontologies aim to model domain conceptualizations so that applications built upon them can be compatible with each other by sharing the same meanings. Life science is one of the most prominent application areas of ontology technology. Many biomedical ontologies have been developed and utilized in real-world systems including Foundational Model of Anatomy (FMA)5, Adult Mouse Anatomy (MA) for anatomy6, National Cancer Institute Thesaurus (NCI)7 for disease and Systematized Nomenclature of Medicine-Clinical Terms (SNOMED-CT)8 for clinical medicine. To integrate and migrate data among applications, it is crucial to first establish mappings between the entities of their respective ontologies. As ontologies in the same domain are often developed for various purposes, there exist several differences in coverage, granularity, naming, structure and many other aspects. It severely impedes the sharing and reuse of ontologies. Therefore, ontology matching (OM) techniques devote to identify mappings across ontologies in order to alleviate above heterogeneities [ 1 ].

In the last ten years, many automatic systems are developed so as to discover mappings between independently developed ontologies and obtain encouraging results (see [ 2, 3 ] for a comprehensive and up-to-date survey). Up to now, the mainstream methods (e.g., LogMap [ 4 ], AML [ 5 ], FCA-Map [ 6 ], XMap [ 7 ]) still focus on engineering features from terminological, structural, extensional (individuals of concepts) information and external resource [ 1 ]. These features are utilized to compute the similarities of ontological entities (i.e., concepts, properties, individuals) for guiding the ontology matching. With the development of deep learning [ 8 ], there also exist several works (e.g., ERSOM [ 9 ], DeepAlignment [ 10 ], SCBOW + DAE(O) [ 11 ] OntoEmma [ 12 ]) that try to shift from feature engineering to representation learning. The assumption is that semantically similar or related words appear in similar contexts. Therefore, word vectors own the potentials that can bring significant value to OM given the fact that a great deal of ontological information comes in textual form [ 10 ]. Nevertheless, there still exist two challenges that need to be solved: – Sparsity Problem for Embedding Learning: One of the main difficulties for embedding learning is of poor capability of encoding sparse entities. Even in largescale medical ontologies with lots of relations, most knowledge graph embedding techniques (e.g., TransE [ 13 ]) are still not applicable. Zhang et al. [ 14 ] observed that the prediction results of entities were highly related to their frequency, and the results of sparse entities were much worse than those of frequent ones. – Limitation Problem for External Resource: Thesaurus is one kind of external resource that is usually employed in matching systems such as WordNet [ 15 ], UMLS Metathesaurus9. In addition, textual descriptions can also be employed for ontology matching [ 11, 12 ]. Nevertheless, these methods based on representation learning rely heavily on the performance of pre-training. Therefore, it may limit their scalability if there exist no dictionaries or sufficient textual corpora.

To address above problems, we propose MultiOM, an alternative ontology matching framework based on embedding techniques from multiple views. The underlying idea is to divide the process of OM into different modules (i.e., lexical-based module, structural-based module, resource-based module) and employ embedding techniques to soften these modules. Existing works [ 16, 17 ] show that identifying multiple views can sufficiently represent the data and improve the accuracy and robustness of corresponding tasks. Therefore, we employ this idea to characterize the process of OM and try to alleviate the sparsity problem for embedding learning indirectly. More precisely, different loss functions are designed based on cross-entropy to model different views 6 http://informatics.jax.org/vocab/gxd/ma ontology 7 https://ncit.nci.nih.gov/ 8 http://www.snomed.org/snomed-ct/ 9 https://www.nlm.nih.gov/research/umls/knowledge sources/metathesaurus/index.html among ontologies and learn the vector representations of ontological entities. With continuous vector representation, we can obtain more similar concepts and discover more potential mappings among ontologies. Furthermore, we design a novel negative sampling tailored for structural relations (e.g., subclassOf relations, disjointWith relations) asserted in ontologies, which can obtain better vector representations of entities for OM.

The contributions of our study are summarized as follows. – We propose an alternative ontology matching framework with embedding techniques from multiple views, and design loss functions based on cross-entropy to model different views for learning vector representations of ontological entities. – We design a novel negative sampling skill tailored for structural relations asserted in ontologies, which can obtain better vector representations of concepts. – We implement our method and evaluate it on real-world biomedical ontologies.

The preliminary result indicates that MultiOM is competitive with several OAEI top-ranked systems in terms of F1-measure. 2 2.1

Related work

Feature-based methods for biomedical ontology matching

There exist various feature-based strategies applied on the scenarios biomedical ontology matching, including terminological-based features, structural-based features and employing external semantic thesauruses for discovering semantically similar entities.

LogMap [ 4 ] relies on lexical and structural indexes to enhance its scalability. To scale to large ontologies and minimize the number of logical errors in the aligned ontologies, LogMap uses a horn propositional logic representation of the extended hierarchy of each ontology together with all existing mappings and employs Dowling-Gallier algorithm to model propositional horn satisfiability.

AML [ 5 ] is originally developed to tackle the challenges of matching biomedical ontologies. It employs various sophisticated features and domain-specific thesauruses to perform OM. Besides, AML introduces a modularization-based technique to extract the core fragments of the ontologies that contain solely the necessary classes and relations caused by disjoint restrictions, which can repair the incoherent alignments effectively.

FCA-Map [ 6 ] is an ontology matching system based on formal concept analysis (FCA), in which five types of formal contexts are constructed in an incremental way, and their derived concept lattices are used to cluster the commonalities among classes and properties at various lexical and structural levels, respectively.

XMap [ 7 ] is a scalable matching system that implements parallel processing techniques to enable the composition of basic sophisticated features. It also relies on the employment of external resources such as UMLS Metathesarus to improve the performance of ontology matching.

PhenomeNet [ 18 ] exploits an axiom-based approach for aligning phenotype ontologies, which makes use of the PATO ontology and Entity-Quality definition patterns so as to complement several shortcomings of feature-based methods.

Feature-based methods mainly employ crafting features of the data to achieve specific tasks. Unfortunately, these hand-crafted features will be limited for a given task and face the bottleneck of improvement. Cheatham and Hitzler showed that the performance of ontology matching based on such engineered features varies greatly with the domain described by ontologies [ 19 ]. As a complement to feature engineering, continuous vectors representing ontological entities can capture the potential associations among features, which is helpful to discover more mappings among ontologies. 2.2

Representation learning methods for biomedical ontology matching

Representation learning have so far limited impacts on OM, specifically in biomedical ontologies. To the best of our knowledge, only five approaches have explored the use of unsupervised representation learning techniques for ontology matching.

Zhang et al. [ 20 ] is one of the first that investigate the use of word vectors for ontology matching. They align ontologies based on word2vec vectors [ 21 ] trained on Wikipedia. In addition, they use the semantic transformations to complement the lexical information such as names, labels, comments and describing entities. The strategy of entity matching is based on maximum similarity.

Xiang et al. [ 9 ] propose an entity representation learning algorithm based on Stacked Auto-Encoders, called ERSOM. To describe an ontological entity (i.e., concept, property), They design a combination of its ID, labels, comments, structural relations and related individuals. The similarity of entities is computed with a fixed point algorithm. Finally, ERSOM generates an alignment based on the stable marriage strategy.

DeepAlignment [ 10 ] is an unsupervised matching system, which refines pre-trained word vectors aiming at deriving the descriptions of entities for OM. To represent the ontological entities better, the authors represent words by learning their representations and using synonymy and antonymy constraints extracted from general lexical resources and information captured implicitly in ontologies.

SCBOW + DAE(O) [ 11 ] is representation learning framework based on terminological embeddings, in which the retrofitted word vectors are introduced and learned by the domain knowledge encoded in ontologies and semantic lexicons. In addition, SCBOW + DAE(O) incorporates an outlier detection mechanism based on a denoising autoencoder that is shown to improve the performance of alignments.

Wang et al. [ 12 ] propose a neural architecture tailored for biomedical ontology matching called OntoEmma, It can encode a variety of information and derive large amounts of labeled data for training the model. Moreover, they utilize natural language texts associated with entities to further improve the quality of alignments.

However, there exist two limitations for above methods. One is the sparsity problem of structural relations. To avoid the poor capability of encoding sparse relations, above methods prefer terminological-based features to learn word vectors for discovering mappings, but they do not make full use of structural relations in ontologies. The other is that these methods rely heavily on the performance of pre-training, which may limit their scalability if there exist no dictionaries or sufficient textual corpora. 3

Muti-view Embedding for Biomedical Ontology Matching

In the scenario of biomedical ontology matching, matching systems mainly focus on mappings of concepts with equivalent relations (Ci; Cj ; ; n). Thus, in the remainder of the paper, we only consider these type of mapping for biomedical ontology matching. 3.1

MultiOM

Existing works [ 16, 17 ] show that identifying multiple views that can sufficiently represent the data and improve the accuracy and robustness of corresponding tasks. Inspired by their works, we characterize the process of OM from multiple views and try to alleviate the sparsity problem for embedding learning indirectly.

O1 O2 Resource-view

Embedding Structural-view

Embedding Lexical-view Embedding

Muti-view Combination

O2

O1

The framework of MutiOM is shown in Fig. 1. Given two biomedical ontologies O1 and O2, we first extract the concepts and their information of ontologies. Then, we divide the process of OM into three embedding modules from different views, which compose of lexical-view embedding, structural-view embedding and resource-view embedding. Domain ontologies in the resource-based module, are treated as bridges to connect source ontology and target one for discovering more potential mappings. With a designed combination strategy based on mutual assessment, we obtain a final alignment among given ontologies.

Different from feature-based methods, we utilize ontological information to learn the continuous vector representations of concepts by embedding techniques, which can discover more potential mappings among ontologies. There exist different granularity of vector representations of modules in MultiOM. In lexical-based module, each concept is divided into several tokens ft1; t2; :::; tng that are represented as k-dimensional continuous vectors ti; tj 2 Rk. The similarity of concepts is measured based on these word vectors by the designed algorithm. Relatively, for structural-based module and resource-based module, each concept C is represented as a d-dimensional continuous vector C 2 Rd, and their similarities are calculated based on cosine measure. Lexical-view Embedding The lexical-view embedding module is mainly based on TFIDF algorithm, which is one of the most effective string similarity metrics for ontology matching [ 19 ]. According to the assumption of TF-IDF, concepts in one ontology can be represented as a bag of tokens. Then, every concept Ci is regarded as a document and the tokens ft1; t2; :::; tlg of each concept are treated as terms. Inspired by the idea soft TF-IDF [ 19 ], we propose an embedding-based TF-IDF strategy to calculate the similarities of concepts, More precisely, the similarity of each concept pair is calculated according to the similarities of their tokens, which is obtained based on the cosine measure of tokens’ vectors representations rather than the string equivalent of them. The corresponding formula is defined as follows.

Sim(C1; C2) = i=1 X wi arg max cos(t1i; t2j); j (1) where C1 and C2 are concepts from O1 and O2, t1i and t2j are vector representations of tokens t1i and t2i that belong to C1 and C2. wi is a weight of token t1i in C1 that is where n is the number of tokens, TFIDF( ) returns the TF-IDF value of each token.

As cosine measure of t1i and t2j is a continuous value, so this embedding-based TF-IDF strategy is able to obtain more similar concepts and discover more potential mappings. Nevertheless, our softened strategy depends on the quality of embedding of tokens and may generate more wrong mappings. Therefore, we utilize pre-training vectors to cover the tokens of ontologies as soon as possible (see Section 4.2). On the other hand, we employ the mappings generated by other embedding modules to assess the quality of these mappings in lexical-view module (see Section 3.3).

Structural-view Embedding As mentioned before, most proposed methods focus on the terminological-based features to learn word vectors for ontology matching, but they do not make full use of structural relations in ontologies. Relatively, we try to generate mappings from the structural view. To obtain more candidate mappings for training embedding of concepts, we assume that the mappings generated by equivalent strings or their synonym labels are correct, and define a loss function based on cross-entropy to optimize the vector representations of concepts. The loss function is defined as follows. lSE =

X (C1;C2; ;1:0)2M

X (C10;C20; ;1:0)2M0 logfSE (C1; C2) log(1 fSE (C10; C20)); (3) where M is a set of candidate mappings f(C1; C2; ; 1:0)g generated by our assumption, M0 is a set of negative mappings. We employ the negative sampling skill [ 13 ] to generate M0 for training the loss function. For each mapping (Ci; Cj ; ; 1:0) 2 M, we corrupt it and randomly replace Ci or Cj to generate a negative triple (Ci0; Cj ; ; 1:0) or (Ci; Cj 0; ; 1:0). fSE (C1; C2) is a score function defined in Eq. 4 to calculate the score of concept pairs, where C1; C2 2 Rd are d-dimensional continuous vectors of concepts C1 and C2 from different ontologies, jj jj2 is the L2-norm. We hope that fSE (C1; C2) is large if concepts C1 and C2 are similar. calculated as follows.

wi = n P TFIDF(t1l) l=1

; fSE (C1; C2) = 2

1 1 + e(jjC1 C2jj2) : Furthermore, we design a negative sampling skill tailored for structural relations asserted in ontologies (e.g., subclassOf relations, disjointWith ) relations. Unlike the uniform negative sampling method that samples its replacer from all the concepts, we limit the sampling scope to a group of candidates. More precisely, for each mapping (Ci; Cj ; ; 1:0) 2 M, if there exist subclassOf relations (e.g., (Ci0; subclassOf ; Ci) or (Cj 0; subclassOf ; Cj )) asserted in ontologies, we need to exclude this replace case. Relatively, for disjointWith relations (e.g.,(Ci0; disjointWith ; Ci) or (Cj ; disjointWith ; Cj 0)), we need to give the highest priority to these relations for replace cases (see Section 4.2). With these constrains for negative sampling, we can obtain better vector representations of concepts for ontology matching.

Resource-view Embedding Inspired by the work in [ 22 ], we consider external ontology as a bridge to connect two concepts from source ontology and target one. We (2) (4) observe that there exist many different yet overlapping biomedical ontologies such as MA—NCI—FMA, FMA—NCI—SNOMED-CT. Compared with textual descriptions or thesaurus, ontologies as external resources can provide some structural assertions, which is helpful to refine the quality of discovered mappings [ 22 ]. Nevertheless, the original idea is mainly based on string equality, which may not discover more similar concepts. Therefore, we employ embedding techniques to soft this framework to discover more potential mappings from this view.

M13 C11

C31 C32

C33 M23

C21

C22 (5) where T is a set of tuples f(C1; C2; C3)g generated by the shared assumption, T 0 is a set of negative tuples that randomly replace C1 or C2. fRE (C1; C2; C3) is a score function defined in Eq. 6 to calculate the score of projected concepts, where C1; C2; C3 2 Rd are d-dimensional continuous vectors of concepts C1, C2, C3 from different ontologies, M13 and M23 are two matrices that project C1, C2 into the embedding space of O3, respectively. We hope that the similar concepts will be projected near their shared concept. Conversely, there existed a semantic distance between dissimilar ones. fRE (C1; C2; C3) = 2

1 1 + e(jjC1 M13 C3jj2+jjC2 M23 C3jj2) : (6)

In order to train two matrices better, we maintain all the vectors of concepts in O3 unchanged and only update the parameters of matrices and concepts from O1 and O2. Furthermore, we take advantage of structural relations in O3 to pre-train the vector representations of concepts, which can adjust semantic distances of concept vectors. As existing KG embedding models face the sparsity problem, we design a loss function based on cross-entropy to achieve this goal that is defined as follows.

lP T =

X (C31;r;C32)2R

X

1 (C301;r;C302)2R0 1 + e(jjC31 C32jj2 logfr(C31; C32) log(1 fr(C301; C302)); (7) where R is a set of relation assertions, involving f(C31; subClassOf ; C32)g [ (C31; PartOf ; C32)g, R0 is a set of negative ones that randomly replace C31 or C32. fr(C31; C32) is a score function that measures the score of (C31; r; C32), C31 and C32 are vector representations of concepts C31 and C32. Notice that, subClassOf and PartOf are not equivalent relations, so we utilize a hyper-parameter to controls the semantic distances of concept vectors. After obtained mappings from different modules, we need to combine them together. A straightforward strategy is collecting all the mappings from these modules and filtering out them with one threshold or stable marriage algorithm. Although this strategy can obtain a high recall in the final alignment, it may also introduce lots of wrong mappings and miss n:m cases about mappings. Therefore, we propose a combination strategy based on mutual assessment.

For convenience, we use OM-L, OM-S, OM-R to represent the alignments generated by lexical-based module, structural-based module, resource-based module, respectively. The concrete procedures are achieved as follows.

Step 1 Merge the mappings from OM-S and OM-R. Their merged result is labeled as OM-SR, in which the similarity of each mapping is selected the large one between OM-S and OM-R.

Step 2 Select the “reliable” mappings of OM-L and OM-SR based on the corresponding thresholds 1 and 2.

Step 3 Assess these “reliable” mappings from OM-L and OM-SR mutually. For example, if one “reliable” mapping belongs to OM-L and its similarity in OM-SR is lower than threshold 3, then we need to remove it. Relatively, if one “reliable” mapping belongs to OM-SR and its similarity in OM-L is lower than threshold 4, then this mapping will be removed.

Step 4 Merge assessed mappings from OM-L and OM-SR and generate a final alignment. For each mapping appearing in OM-L and OM-SR at the same time, its similarity is selected the large one. 4

Experiments

To verify the effectiveness of MultiOM, we used Python to implement our approach with the aid of TensorFlow10 and parse ontologies by OWLAPI11. The experiments were conducted on a personal workstation with an Intel Xeon E5-2630 V4 CPU which has 64GB memory and TiTAN XP GPU. Our approach12 can be downloaded together with the datasets and results. 10 https://www.tensorflow.org/ 11 http://owlapi.sourceforge.net/ 12 https://github.com/chunyedxx/MultiOM 4.1

Datasets

We collect the biomedical ontologies from Anatomy Track and Large BioMed Track in OAEI13 (Ontology Alignment Evaluation Initiative), which is an annual campaign for evaluating ontology matching systems that attracts many participants all over the world. Furthermore, this campaign provides uniform test cases and standard alignments for measuring precision, recall and F1-measure for all participating systems. 4.2

Experiment Settings

We select several strategies to construct the baseline systems to verify the effectiveness of our model. The following is the detail construction of strategies in our experiments. – StringEquiv: It is a string matcher based on string equality applied on local names of entities. – StringEquiv + Normalization (StringEquiv-N): It employs normalization techniques before execute StringEquiv matcher. – StringEquiv + Synonym (StringEquiv-S): It extends the synonym of concepts when executing the StringEquiv matcher. – StringEquiv + Synonym + Reference Ontology (StringEquiv-SR): It introduces external ontologies as bridges to connect concepts based on StringEquiv-S. – StringEquiv + Synonym + Normalization (StringEquiv-NS): It extends the synonym of concepts when executing the StringEquiv-N. – StringEquiv + Normalization+ Synonym + Reference Ontology (StringEquiv-NSR): employs normalization techniques before execute StringEquiv-SR.

For MultiOM, we use stochastic gradient descent (SGD) as an optimizer and the configuration of hyper-parameters is listed below: Dimensions of concepts and matrices are set to d=f50; 100g and dM=f50; 100g. The mini-batch size of SGD is set to Nbatch=f5; 10; 20; 50g. We select the learning rate among f0.01; 0:02; 0:001g and f1; 3; 5; 10g negative triples sampled for each positive triple. The whole training spent 1000 epochs. In lexical-based module, the vector presentations of tokens mainly come from the linkage14 of the work [ 11 ], whose dimension is set to 200. For some tokens without vector presentations, we initialize them randomly and enforce constrains as jjt1ijj2 1 and jjt2jjj2 1. In resource-view embedding module, we employ TransE [ 13 ], ConvE [ 23 ] and pre-training function 7 to initialize the vector representations of concepts in external ontologies. is set to f0.01; 0:05; 0:10g in function 7 for controlling the semantic distances of concept vectors. For negative sampling strategy, we collect all the related structural assertions of concepts. When one concept is selected as a replacer, we retrieve the structural assertions of this concept and execute the replacement based on its relations with the original concept. During this process of replacement, disjointWith relations own the highest priority and subclassOf relations should be excluded. Finally, the result of MultiOM is generated by the combination strategy, and we set the related thresholds 1 = 0:8, 2 = 0:95, 3 = 0:65, 4 = 0:3.

In order to show the effect of our proposed negative sampling, a symbol “-” added to the symbol represented module (or merged one) indicates that this module is not equipped with negative sampling tailored for structural relations.

Evaluation Results

Table 1 lists the matching results of MultiOM compared with baseline systems. We observe that merging more strategies can improve the number of mappings. Although it slightly decreases the precision of alignments, it can increase the recall and F1-measure as a whole. Relatively, MultiOM further improves the recall and F1-measure of alignments because continue vector representations of concepts can obtain more similar concepts and discover more potential mappings. Moreover, the performance of MultiOM is better than MultiOM in term of F1-measure. The main reason is that employing structural relations are helpful to distinguish the vector representations of concepts.

Table 2 shows the results of different combination with embedding-view modules. Overall, merge more embedding modules, the performances of alignments are better. For lexical-view module, softened TF-IDF (denoted as OM-L) is better than original TF-IDF in terms of F1-measure because continuous vectors representing tokens can provide more semantic information than single strings for calculating the similarity of concepts. For resource-view embedding module (denoted as OM-R), ConvE and our pre-training function are better than random initialization because both of them can utilize structural relations to adopt vector representations of concepts in the semantic space. Nevertheless, compared with 20 minutes spent in function 7, ConvE took nearly 24 hours to obtain the vector presentations of concepts. Notice that, it is not suitable for TransE to pre-train the vector presentations of concepts. We analyze that sparse structural relations of ontologies and its simplified score function limit its capability. Overall, we observe that employing new negative sampling strategy in embedding-view modules (i.e., OM-S, OM-RS, MultiOM) is helpful to improve the quality of alignments further in terms of precision and F1-measure.

Table 3 lists the comparison of MultiOM with OAEI 2018 top-ranked systems based on feature engineering and representation learning. Preliminary result shows that MultiOM can be competitive with several promising matching systems (e,g, FCAMapX, XMap) in terms of F1-measure. Nevertheless, there still exists a gap compared with the best systems (e.g., AML, SCBOW + DAE (O)). We analyze that lexical-based module and simplified combination strategy may become the main bottlenecks of MultiOM. Benefited from thesauruses (e.g., UMLS) and optimized combination strategy, most top-ranked systems can obtain better performances of OM tasks. In addition, most systems (e.g., AML, LogMap) employ alignment debugging techniques, which is helpful to improve the quality of alignment further. But we do not employ these techniques in the current version. We leave these issues in our future work.

Conclusion and future work

In this paper, we presented an alternative OM framework called MultiOM, in which different loss functions were designed based on cross-entropy to model different views among ontologies and learn the vector representations of concepts. We further proposed a novel negative sampling skill tailored for structural relations, which could obtain better vector representations of concepts. We implemented our method and evaluated it on real-world biomedical ontologies. The preliminary result indicated that MultiOM was competitive with several OAEI top-ranked systems in terms of F1-measure.

In the future work, we will explore following research directions: (1) As candidate mappings and tuples are not enough, we will extend MultiOM to an iterative framework. (2) Recently, Zhang et al. [ 17 ] presented combination strategies for entity alignment based on embedding techniques. Incorporating these combination strategies into MultiOM may facilitate improving the quality of mappings. (3) Senior symbolic reasoning techniques (e.g., incoherent checking) could be served for training process and alignment generation. We will merge them into MultiOM for improving its performances.

References

1. Je´roˆme Euzenat and Pavel Shvaiko . Ontology Matching . Springer Science, 2013 .

2. Lorena Otero-Cerdeira , Francisco J Rodr´ ıguez-Mart´ınez, and Alma Go´mez-Rodr´ıguez. Ontology matching: A literature review . Expert Syst. Appl. , 42 ( 2 ): 949 - 971 , 2015 .

3. Ian Harrow , Ernesto Jime´ nez -Ruiz, Andrea Splendiani, Martin Romacker, Peter Woollard , Scott Markel, Yasmin Alam-Faruque, Martin Koch, James Malone, and Arild Waaler . Matching disease and phenotype ontologies in the ontology alignment evaluation initiative . Journal of Biomedical Semantics , 8 ( 1 ): 1 - 13 , 2017 .

4. Ernesto Jime ´nez-Ruiz and Bernardo Cuenca Grau . Logmap: Logic-Based and Scalable Ontology Matching . In ISWC , pages 273 - 288 , 2011 .

5. Daniel Faria , Catia Pesquita, Emanuel Santos, Matteo Palmonari, Isabel F Cruz, and Francisco M Couto. The AgreementMakerLight Ontology Matching System . In OTM Conferences , pages 527 - 541 , 2013 .

6. Mengyi Zhao , Songmao Zhang, Weizhuo Li , and Guowei Chen . Matching biomedical ontologies based on formal concept analysis . Journal of Biomedical Semantics , 9 ( 1 ): 11 , 2018 .

7. Warith Eddine Djeddi and Mohamed Tarek Khadir. A Novel Approach Using Context-Based Measure for Matching Large Scale Ontologies . In DaWaK, pages 320 - 331 , 2014 .

8. Yann LeCun , Yoshua Bengio, and Geoffrey Hinton . Deep learning . Nature , 521 : 436 , 2015 .

9. Chuncheng Xiang , Tingsong Jiang, Baobao Chang , and Zhifang Sui . ERSOM: A Structural Ontology Matching Approach Using Automatically Learned Entity Representation . In EMNLP , pages 2419 - 2429 , 2015 .

10. Prodromos Kolyvakis , Alexandros Kalousis, and Dimitris Kiritsis. DeepAlignment: Unsupervised Ontology Matching with Refined Word Vectors . In NAACL , pages 787 - 798 , 2018 .

11. Prodromos Kolyvakis , Alexandros Kalousis, Barry Smith , and Dimitris Kiritsis . Journal of Biomedical Semantics , 9 ( 1 ): 21 , 2018 .

12. Lucy Wang , Chandra Bhagavatula, Mark Neumann, Kyle Lo, Chris Wilhelm, and Waleed Ammar . Ontology alignment in the biomedical domain using entity definitions and context . In BioNLP , pages 47 - 55 , 2018 .

13. Antoine Bordes , Nicolas Usunier, Alberto Garcia-Duran, Jason Weston , and Oksana Yakhnenko . Translating Embeddings for Modeling Multi-relational Data . In NeurIPS , pages 2787 - 2795 , 2013 .

14. Wen Zhang , Bibek Paudel, Liang Wang, Jiaoyan Chen, Hai Zhu, Wei Zhang, Abraham Bernstein, and Huajun Chen . Iteratively learning embeddings and rules for knowledge graph reasoning . In WWW , pages 2366 - 2377 , 2019 .

15. George A Miller. WordNet: A Lexical Database for English . Communications of the ACM , 38 ( 11 ): 39 - 41 , 1995 .

16. Meng Qu , Jian Tang, Jingbo Shang, Xiang Ren, Ming Zhang, and Jiawei Han. An attentionbased collaboration framework for multi-view network representation learning . In CIKM , pages 1767 - 1776 , 2017 .

17. Qingheng Zhang , Zequn Sun, Wei Hu, Muhao Chen, Lingbing Guo, and Yuzhong Qu . Multiview Knowledge Graph Embedding for Entity Alignment . In Proceedings of IJCAI , 2019 .

18. Miguel A´ ngel Rodr´ıguez-Garc´ıa, Georgios V Gkoutos, Paul N Schofield, and Robert Hoehndorf. Integrating phenotype ontologies with PhenomeNET . Journal of Biomedical Semantics , 8 ( 1 ): 58 , 2017 .

19. Michelle Cheatham and Pascal Hitzler . String Similarity Metrics for Ontology Alignment . In ISWC , pages 294 - 309 , 2013 .

20. Yuanzhe Zhang , Xuepeng Wang, Siwei Lai, Shizhu He, Kang Liu, Jun Zhao , and Xueqiang Lv . Ontology Matching with Word Embeddings . In CCL, pages 34 - 45 . 2014 .

21. Tomas Mikolov , Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean . Distributed Representations of Words and Phrases and their Compositionality . In NeurIPS , pages 3111 - 3119 , 2013 .

22. Songmao Zhang and Olivier Bodenreider . Experience in Aligning Anatomical Ontologies . International Journal on Semantic Web and Information Systems , 3 ( 2 ): 1 - 26 , 2007 .

23. Tim Dettmers , Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel . Convolutional 2D Knowledge Graph Embeddings . In AAAI , pages 1811 - 1818 , 2018 .