KGMatcher is a scalable and domain independent matching tool that matches the schema (classes) of larger 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 stringbased matcher. This year is the rst OAEI participation of KGMatcher, and it is the best performing system on the common knowledge graph track. Although KGMatcher results are promising, further improvements of the matching techniques' and matcher combination can be introduced.
1.1
Combining di erent matching techniques is a common practice in ontology
matching tools. Matching techniques are divided into three main categories [
The matcher proposed here is a hybrid approach that combines an element
level matcher to an instance-based one. The rst matcher uses entity labels to
generate candidate pairs, and the latter produces candidate class alignments
by only using annotated instance names. While conventional ontologies mainly
focus on modelling classes and properties, Knowledge Graphs (KGs), particularly
those available on the web, are large-scale and describes numerous instances [
KG classi er can then be used to classify any instance name into one of the classes
from that KG. This approach is domain independent and is capable of coping
with KGs with unbalanced populations (for details, see [
Given two input knowledge graphs, O and O0, where O contains a set of classes O = fC0 ; C1 ; ::; COig, and each class contains a set of instances COi = fe0; ei1; :::; eing. i SimilarlOy, OO0contains a set of classes such that O0 = fCO00 ; CO10 ; ::; COj0 g where CO0 = fej0; ej1; :::; ejmg. KGMatcher has two main components: an instancej based matcher and a name matcher. The work ow of KGMatcher is illustrated in gure 1. Preprocessing 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 dictionary, the method is able to infer the spaces between words and replace them with a space. Instance-based Matcher The rst 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 classi cation method where a KG classi er is trained using one KG's instances data. Later on, that classi er is used to classify any instance name into one of the classes from that KG.
{ Exact Name Filtering. The matcher starts by applying an exact name lter
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 which reduces
the search space of the instance-based matcher.
{ Undersampling. As the instance distribution in KGs tends to be highly
imbalanced, the goal of this step is to balance the number of populated instances
in the input KGs to avoid biased classi cation results. KGMatcher follows a
resampling strategy aimed at undersampling classes with instances number
exceeding the average number of instances per class in that KG, i.e., majority
classes. The standard TF-IDF [
Instance Matching 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 lowercasing, 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. 1.3
KGMatcher is mainly developed with Python. To facilitate reusing and
evaluating KGMatcher and for the OAEI submission, KGMatcher was packaged using
a SEALS client. The wrapping services 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 le. The results include the following OAEI tracks: Conference, Knowledge Graph, and Common Knowledge Graphs track. 2.1
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 justi es the negative impact of the undiscovered property alignments on the matcher performance on this task. Further, given that the Conference track datasets do not include enough number of instances to apply the instance-based matcher, the name matcher is the only matcher applied to map classes. In terms of the new experimental cross-domain test case of mapping DBpedia and OntoFram, KGMatcher performance (0.55) better than both baselines StringEquiv and edna which have the scores 0.42 and 0.45 respectively. 2.2
In the Knowledge Graph track, KGMatcher was able to generate results for all the 5 test cases at both classes and instances level only. 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.
Along with other 6 matchers, KGMatcher was able to complete the task of matching the classes from two cross-domain and common KGs. KGMatcher obtained a precision of 0.97 and a recall of 0.91. With an F1 score of 0.94, KGMatcher is the best performing matcher on this track. 3
The results of KGMatcher has 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' combination. It is important to note that the
performance of KGMatcher instance-based component depends on the dataset
nature. Since KGMatcher is learning KG classi ers by using general pretrained
word embedding models, the more representative the KG instances of real-world
entities, the better are the instance classi cation results. Figure 2 shows the
different 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 [
In future versions of KGMatcher, we would like to improve the combination of techniques used within the name matcher. Currently, this component is rather simple and unable to discover matching pairs with high lexical complexity. This has also a ected the matcher's performance on datasets where instance data are not existing or di cult to classify. Additionally, improving the instance-based matcher by further studying other sampling approaches and experimenting with other machine learning methods, would likely improve the overall performance of the matcher. 4
As part of OAEI 2021, this paper presents KGMatcher, a matching tool that utilizes instance data annotated within large KGs to map their classes. The process is done by learning KG classi ers, which are able to classify instances into a particular KG class. The results suggest that a hybrid approach that incorporate an instance-based technique can be highly e ective for matching large cross-domain KGs.