DBpediaとドメイン固有のナレッジグラフとのマッチングに関する予備的研究

抄録

This paper shows an approach to integrating general-purpose Knowledge Graphs (KGs) and domain-speci_c KGs. General purpose KGs cover di_erent topics but lack speci_c topics or details. On the other hand, KGs extracted from speci_c domains usually represent only a small set of entities compared to the general purpose KG, while they represent entities from their domain with more details. So they can complement the information if used together. Current matching approaches are tested on datasets with one-to-one assumption or relatively small instance sizes. This research explored matching KG extracted from speci_c communities and DBpedia as a general-purpose KG in a real-world case. We used a traditional matching algorithm (PARIS) with a BERT model to _lter the result and random walk to expand the possible matches. Firstly, we executed PARIS for the entire KG and selected the obtained matches above a threshold. Secondly, the algorithm embedded the abstract of the entities using the BERT model, calculated the similarity between the vectors, and _ltered the matches. In the last step, the algorithm used the _ltered matches as seeds to apply random walks and created a sub-graph for each KG. Then, the instances of the sub-graphs were matched using string similarity between the labels and similarity between the abstracts using BERT when available on both sides. We tested the proposed approach between the entire DBpedia and our KG and improved the obtained matches. We found that the generated matches contained many entities with de_cient information in DBpedia, so the matching process can be used to identify and complement those entities.