Ontology Matching Method Based on Gated Graph Attention Model

1  Introduction

In the development of the Semantic Web [1], the study of ontology has a significant impact on its advancement. Currently, ontology technologies are extensively utilized across various domains, such as biomedicine, finance, engineering, law, and cultural heritage [2]. However, in the field of ontology engineering, different ontologies may use different representations and structures when describing the same concepts [3,4], which can cause a decrease in data sharing and system interoperability, and increase the complexity of knowledge integration and data fusion. This issue is referred to as ontology heterogeneity. To solve these issues, researchers have proposed the ontology matching approach [5], which aims to identify corresponding semantically identical or similar concepts across different ontologies.

In recent years, Representation Learning has become a main approach to solving problems in various domains, with models capable of automatically extracting features or representations of features [6]. Consequently, we focus on the research of ontology matching methods from the perspective of representation learning methods. Existing representation learning methods still have the following three problems in acquiring and fusing semantic and structural features of ontologies: (1) In terms of semantic embedding, existing word embedding methods such as Word2Vec [7] and GloVe [8], primarily used for natural language processing (NLP) tasks, learn the vector space representation of words by analyzing the co-occurrence of vocabulary in textual data. These methods excel in handling natural language text, but they are not directly applicable to Web Ontology Language (OWL) ontologies, as the structure and semantic characteristics of OWL ontologies significantly differ from natural language text. To address this issue, the OWL2Vec* [9] word embedding technique has been proposed, which transforms complex ontology logic and structural information into vector representations that are easily processed by machines, thereby efficiently encoding the semantic information of OWL-formatted ontologies. Its working principle involves converting OWL ontologies into graphs and then performing random walks on these graphs to generate structural documents that serve as the input corpus. Subsequently, it combines the graph structure of the ontology, logical constructors, and lexical information from textual annotations to create a comprehensive corpus. Finally, by training a Word2Vec model on the corpus, it generates embedding representations that capture the semantic information of concepts within the ontology. Although OWL2Vec takes into account the ontology’s vocabulary and logical constructors, it does not fully consider the true meaning of concepts themselves and their hierarchical relationships when encoding the semantic information of the ontology. (2) In terms of structural embedding, the Graph Convolutional Network (GCN) model can capture and utilize the graph structure information of the ontology when extracting the graph node information of the ontology [10]. The basic idea is to update the embedding representation of the current node by aggregating the characteristics of the neighboring node. However, it struggles to model and filter the importance of distant nodes associated with a central node. Furthermore, the GCN model needs to update the whole graph when updating node features, and the fusion is less efficient when there are more neighboring nodes. (3) When integrating ontological semantic and structural embedding representations, the existing direct concatenation strategies [11], while simple, fail to adequately consider the importance of semantic and structural features in the matching task.

To address the above three problems, we propose corresponding solutions from different knowledge perspectives, and the main contributions are as follows:

(1) We propose an Ontology Matching Method Based on a Gated Graph Attention Model (OM-GGAT). Firstly, in terms of semantic embedding, to enrich the contextual semantics of concepts, we employ the OWL2Vec* method to encode the semantic knowledge related to concepts in the ontology into vectors. Additionally, by embedding the path information of concepts (i.e., the complete path from the root node to the concept), we can better understand the true meaning of the concepts themselves and their hierarchical relationships with each other. Secondly, in terms of structural embedding, the ontology is transformed into the corresponding graph structure based on semantic relationships. When extracting features of the ontology graph nodes, we utilize the Graph Attention (GAT) model to assign different attention weights to each neighboring node of the central concept, adaptively aggregating the features of neighboring nodes and capturing key distant node information. Finally, a gated network is designed to effectively integrate semantic and structural embedding representations, achieving more accurate ontology matching.

(2) Unsupervised and semi-supervised ontology matching task comparison experiments on three public datasets demonstrate that OM-GGAT can effectively improve ontology matching efficiency and provide a solution for knowledge fusion. Separate experiments on semantic and structural approaches are also conducted to demonstrate that the OM-GGAT approach is effective in considering both semantic and structural approaches on the ontology matching task.

2  Related Work

In ontology, concepts are represented by vectors constructed by feature engineering, and the ontology matching task can be transformed into the vector-based similarity calculation task between different ontologies. For example, Kolyvakis et al. [12] proposed an ontology-matching framework that uses word embedding technology to capture semantic similarity between ontologies. The real situation of semantic similarity of concepts can be distinguished by connecting the context semantics of concepts. Chen et al. [9] proposed an OWL2Vec* method that encodes semantic information in OWL ontologies by combining random walk algorithms with word embedding techniques. Xue et al. [13] proposed an Ontology Meta-Matching technique (OMM) based on deep reinforcement learning. This technology integrates multiple similarity measurement methods to discover heterogeneous entities between different ontologies. Li et al. [14] proposed a knowledge representation learning model TransO based on constraint concept types, relations, and hierarchical information, which can effectively model relations, complete the reasoning of knowledge graphs, and maintain low model complexity. However, ontology not only contains rich semantic features, but also its structural information can reveal the interrelation between concepts. Sentürk et al. [15] proposed a graph-based ontology matching framework, which converts ontology concepts and relationships into graph structures and uses subgraph mining technology to carry out effective ontology matching. These methods only rely on a single feature of the ontology, whether it is a semantic feature or structural feature, which may lead to the loss of knowledge in the matching process.

Existing research points out that it is difficult to accurately judge whether two concepts match each other using only one similarity measurement method, but combining multiple similarity strategies can significantly improve the accuracy of matching. For example, Duan et al. [11] proposed an ontology-matching method based on word embedding and structural similarity. The method mainly distinguishes semantic similarity and description association by improving word vectors. Then the ontology is transformed into a graph and the SimRank algorithm is used to calculate the structural similarity, to realize the one-to-many matching task. In recent years, researchers have begun to use Machine Learning (ML) methods to effectively integrate a variety of similar ontology matching methods. For example, Efeoglu et al. [16] proposed a Graphmatcher ontology matching system based on Graph Representation Learning. It uses Graph Representation Learning methods and Graph Attention Mechanism to compute high-level representations of classes and their related terms in ontologies to identify and align semantically similar entities in different ontologies. Xue et al. [17] proposed an ontology matching method based on an Interactive Compact Genetic Algorithm (ICGA). This method uses a compact coding mechanism and expert interaction mechanism to improve the performance and alignment quality of the algorithm. He et al. [18] proposed an ontology alignment method based on Bidirectional Encoder Representations from Transformers (BERT). In addition, to address the limitations of current suboptimal reference mappings and limited support for evaluation of machine learning-based systems, He et al. [7] proposed the DeepOnto method and an Ontology Pruning method, which could improve the relative integrity of reference mapping.

At present, many ontology matching studies focus on using information such as concept name, ontology structure, and external resources to improve matching tasks [19], however, these methods often do not take into account the deep semantic information of concepts, which leads to low accuracy of matching. In addition, although semantic-based matching methods improve accuracy by extracting semantic information, they still fail to make full use of the intrinsic logical structure of ontology to reveal the potential semantic connections between concepts. Furthermore, future research should investigate more efficient fusion strategies to integrate diverse similarity information, enhancing matching performance.

3  Ontology Matching Method Based on Gated Graph Attention Model

3.1 Related Concepts

Since we only focus on the matching relationship between concepts in ontology. The ontology [20] is formally defined as the triplet form in Eq. (1).

O=<C,P,H>(1)

where O represents ontology, C represents the set of Classes in the ontology, P represents the set of data properties and object properties in the ontology, H represents the Hierarchical Relationships of classes.

In order to ensure the uniformity of terminology, classes, and property are collectively referred to as concepts. When the structure of an ontology is more complex, the OWL language is commonly used for its description. Therefore, in the experimental section, we will process ontologies in OWL format. Given source ontology O1 and target ontology O2, the Ontology matching [21,22] task can be formally defined as a quadruple form of Eq. (2).

MR=<C1,C2,R,f(C1,C2,R)>(2)

where MR indicates the matching result of concepts between ontologies. C1∈O1 represents the concept in source ontology O1. C2∈O2 represents the concept in target ontology O2. R describes the relationship between the concepts C1 and C2 (including equal, unequal, inclusive, intersecting, etc.). f(C1,C2,R) represents the confidence of the relationship between concepts C1 and C2 calculated using the matching method, and f has a value interval of [0, 1]. The greater the similarity value, the higher the probability that the concepts C1 and C2 represent the same thing.

3.2 Model Implementation

3.2.1 Model Overview

Applying the concept of attention mechanisms to ontology matching methods can further enrich the semantic knowledge of the ontology and capture the importance of neighboring nodes between different concepts. Therefore, we propose an ontology matching method based on the Gated Graph Attention Model. The model diagram is shown in Fig. 1.

Figure 1: Overview of OM-GGAT

The model mainly includes the following four steps: (1) Semantic Embedding: Using OWL2Vec* method to encode the semantic knowledge of the ontology and embed the related path information of the concept. (2) Structural Embedding: The GAT model is used to extract node features of concepts. (3) Gated Network: Gating networks are introduced to effectively integrate the embedding of semantics and structure. (4) Ontology Matching Process: Calculate the similarity between concepts and output the final matching results in the form of matching pairs.

3.2.2 Semantic Embedding

In order to better understand the true meaning of concepts in an ontology and the relationship between concepts, we embedded path information related to concepts. Specific steps are described as follows.

First, in the OWL ontology, the OWL2Vec* method is used to encode the semantic knowledge related to concepts into vectors. Then, the related path information of the concept is embedded, and the path refers to the sequence of concepts obtained from the “root” node of the ontology to the concept node in turn. For example, for the concept node microcephaly in the DOID ontology, the complete path information from the root node owl: Thing to the concept node microcephaly is “Thing/disease/physical disorder/microcephaly”. Therefore, the path sequence for the concept node microcephaly can be represented as (Thing, disease, physical disorder, microcephaly). The path sequence is formally described as path1=(C11,C12,…,C1n) for the given concepts C1, and path2=(C21,C22,…,C2m) for the concept C2. Where, n and m represent the length of the path sequence respectively. The embedding method is shown in Eq. (3).

EPsemantic(C1)={[Esemantic(C1)||VC11||VC12…||VC1n]}EPsemantic(C2)={[Esemantic(C2)||VC21||VC22…||VC2m]}(3)

where Esemantic(C1) and Esemantic(C2) represent the semantic embedding of concepts C1 and C2 obtained using the OWL2Vec* method, respectively. EPsemantic(C1) and EPsemantic(C2) represent the semantic embedding of concepts C1 and C2 embedded with corresponding path information, respectively. VC11 and VC21 represent the vector of the first path node in concepts C1 and C2, respectively, and so on. [⋅||⋅] represents the concatenation operation.

3.2.3 Structural Embedding

When extracting features of nodes in an ontology graph, compared to other attention mechanisms, GAT has the advantage of implicitly assigning different weights to different nodes in the neighborhood, which helps to reduce the impact of noise propagation on nodes in the graph and enhances the effectiveness of structural embedding vectors [23]. In addition, the attention score mechanism allows us to capture and filter the importance of distant nodes related to the central node, thereby further improving the performance of the model. Therefore, in order to