KGMatcher+ is a scalable and domain-independent matching system that matches the schema (classes) of large Knowledge Graphs by following a hybrid matching approach. KGMatcher+ is composed of an instance-based matcher which only uses annotated instances of knowledge graph classes to generate candidate class alignments and a string-based matcher. This year is the second OAEI participation of KGMatcher+, formerly known as KGMatcher. More improvements have been added to the matcher, particularly in terms of handling imbalanced class distribution, and it is the best-performing system in the common knowledge graphs track this year.
makes the matcher particularly useful for matching large KGs with numerous populated and
overlapping instances such as DBpedia [
Given two input knowledge graphs, and ′, where has a set of classes = {0, 1 , .., , aonfdclaesascehs csluacshs cthonattain′s=a {set0o′f, in1st′ a,n..c,es′}w=he{re0,1 , ..., },.S1 ,i m...i,l arly},.K′GcMonattcahinesr+ahsaest} two main components: an instance-based matcher and′ =an{am0e matcher. The workflow of KGMatcher+ is illustrated in figure 1.
Given the two input KGs, the matcher starts by parsing and indexing the lexical data of the two KGs separately. Following the standard free text search/index approach, an index is created for each KG where each class is treated as a document and the content of each ‘document‘ is the concatenation of the class’s instance labels. In addition to the standard text cleaning processes, a word segmentation method is applied in order to separate multi-word entities, e.g., academicfield. Using a general dictionary, this method is able to infer the spaces between words and replace them with a space.
The first matching component of KGMatcher+ belongs to the extensional matcher category. It
uses a self-supervised machine learning approach to map KG classes based on their instances
overlap. The matching is done in a two-way classification fashion where a KG classifier is
trained using one KG’s instances data. Later on, that classifier is used to classify any instance
name into one of its classes. Here, we summarize the matching process of KGMatcher+, however,
readers may refer to [
• Exact Name Filtering. The matcher starts by applying an exact name filter to exclude class
labels that exist in both input KGs. Given the large number of classes in typical KGs, this
step works as a blocking strategy that reduces the search space of the instance-based
matcher.
• Resampling KGs Instances. The class distribution in typical public KG tends to be highly
imbalanced. Thus, the goal of this step is to balance the number of populated instances
in the two input KGs to avoid biased classification results. The previous version of
the system was only targeting the majority, i.e., large classes, by undersampling their
instances using TF-IDF approach [
1The parameters selected for the DNN model: an input layer of pre-trained word embedding model followed by four fully connected hidden layers with 128, 128, 64, 32 rectified linear units. A dropout layer of 0.2 is added between each pair of dense layers for regulation. Finally, a softmax layer for multi-class classification, taking the total number of classes in the KG we are training a classifier for.
2The input layer is the Google News token-based model https://tfhub.dev/google/tf2-preview/nnlm-en-dim128/1
• KG1 alignment elicitation is the process of generating candidate pairs using the classifier
trained on the first input KG. Candidate pairs are generated by iteratively applying the
classifier to instances in the other KG’s classes. As a result, each instance name in
′ is now classified into a class in . The candidate pair (,′ ) is added to the first
candidate alignments set →′ if the majority of
. A similarity score between [
• KG2 alignment elicitation is similar to the above-illustrated elicitation process. However,
the roles of the two KGs are reversed where ′ , i.e., the classifier trained on the second
KG (′), is applied to instances in order to obtain the second candidate alignment set
′→.
• Similarity computing is where KGMatcher+ combines the two candidate alignment sets
resulting from the two-way classification method. First, the matcher separately stores
each directional alignment in an alignment matrix of a ||.|′| dimension. The two
matrices are then aggregated into one matrix by taking the average similarity score of
each pair. For example, if (6 ,3′ ,0.88) in →′ and (5 ,3 , 0.64) in ′→) their
′
aggregated similarity value will be 0.76. Consequently, the final alignments for this
matcher are chosen by following the state-of-the-art automatic final alignment selection
approach introduced in [
The second component of KGMatcher+ is an element-level matcher, which measures the similarity of KG class labels. First, the edit distance of each class pair is measured, and then their semantic similarity is measured by a word embedding method. First, the Levenshtein distance is calculated for each class pair. Then, in terms of the word embedding similarity, a pre-trained word2vec model is used to represent class labels before measuring their cosine similarities. The semantic similarity is measured in a Vector Space Model, where words with high semantic relations are often represented closer to each other. In the case of multi-word labels, the vector representation of each word composing the label is aggregated with an element-wise average of the composing word vectors. Finally, the maximum of the two similarity measures is chosen as the name similarity of that pair. The threshold value of the name matcher is set to 0.8. To illustrate, if the word embedding similarity of (RailwayStation,TrainStation) is 0.83 while their Levenshtein distance is 0.56, the maximum similarity value, i.e., the word embedding similarity which is also higher than 0.8. Nonetheless, in case the two similarity scores are lower than the threshold, that pair will be excluded from the candidate alignment.
KGMatcher+ combines the results generated from the two matching components by following the same method described earlier to combine the two instance classification alignments.
For the OAEI participation, we have adapted KGMatcher+ to also match the instances of KGs. The instance matching component is very simple. First, standard text preprocessing techniques such as lower casing, and removing stopwords and non-alphanumeric characters are applied. Then, KGMatcher+ generates candidate instance pairs based on the existence of the label in the opposite knowledge graph.
KGMatcher+ is mainly developed with Python. To facilitate reusing and evaluating KGMatcher+
and for the OAEI submission, it was packaged using a SEALS client. The wrapping service from
the Matching EvaLuation Toolkit (MELT) [
In this section, we present and discuss the results for each of the OAEI tracks where KGMatcher+ was able to produce a non-empty alignment file. The results include the following OAEI tracks: Conference, Knowledge Graph, and Common Knowledge Graphs track.
In the Conference track, when following the rar2-M3 evaluation, KGMatcher+ F1 score (0.52) is slightly lower than both baselines, i.e., StringEquiv (0.53) and edna (0.56). This particular evaluation, i.e., M3 takes into consideration both class and property matches. The fact that KGMatcher+ does not match property justifies the negative impact of the undiscovered property alignments on the matcher’s performance on this task. Further, given that the Conference track datasets do not include enough instances to apply the instance-based matcher, the name matcher is the only matcher applied to map classes. In terms of the cross-domain test case of mapping DBpedia and OntoFram, KGMatcher+ is the second to the best-performing system.
In the Knowledge Graph track, KGMatcher+ was able to generate results for all 5 test cases at both classes and instances level. In terms of class matching, the matcher yields satisfactory results, with 0.79 for F1 score. The added instance matcher has positively impacted the overall matcher result on this task, with a precision of 0.94, a recall of 0.66, and F1 of 0.82. KGMctcher+ is the second to the best-performing system in this track.
In this track, KGMatcher+ was able to complete the task of matching the classes from 4 crossdomain and common KGs. On the task of matching NELL and DBpedia, the matcher obtained the highest F1 score of 0.95. In terms of the second task, which maps classes from Yago and Wikidata, KGMatcher+ is also the best-performing matching system with a recall of 0.83 and an F1 score of 0.91. KGMctcher+ yields the best performance results on this track.
The results of KGMatcher+ have been very encouraging. In the common knowledge graph track,
it achieves outstanding results. This indicates that our hybrid approach, utilizing instances data
to map KG classes, is able to outperform systems that use other matchers’ combinations. It is
important to note that the performance of KGMatcher+ instance-based component depends on
the dataset nature. Since KGMatcher+ is learning KG classifiers by using general pre-trained
word embedding models, the more representative the KG instances of real-world entities, the
better the instance classification results. Figure 2 shows the diference between the performance
when classifying instances from common KGs, e.g., NELL, compared to a single domain KG from
the knowledge graph track. Note that the latter mainly annotates classes in the entertainment
domain [
(a) (b)
As part of OAEI 2022, this paper presents KGMatcher+, a system for matching the schema of large-scale and domain-independent KGs by utilizing instances data. The process is done by learning KG classifiers, which are able to classify instances into a particular KG class. The results suggest that a hybrid approach that incorporates an instance-based technique can be highly efective particularly for matching large cross-domain KGs with imbalanced class distribution, such as Yago and Wikidata.