Enhancing Panoramic Competency Through Link Prediction in Question Knowledge Graphs using a Language Representation Model

Recently, panoramic knowledge has been required. On the other hand, multiple choice questions are suitable for efficient self-learning. Therefore, the purpose of this study is to create multiple choice questions that can reinforce learners' panoramic knowledge. Specifically, we proposed a method for automatically generating multiple choice questions that use Linked Data to present relevant information to give respondents an overall picture of relevant knowledge. There is some research on the methods that generated questions by extracting small subgraphs from the knowledge graphs consisting of entities(words) and relations(links) between the entities and hiding target words (correct answer words). In this study, our goal is to enhance the panoramic of the subgraphs of a specified size by using the link prediction method to complement edges and represent relationships not present in the knowledge graph when generating questions targeted at specific fields. The method of complementing edges involves first inputting two words as subject and object in Knowledge Graph to calculate the cosine similarity using a pretrained language model based on Wikipedia and Wikidata, then predicting the links as a predicate that should be complemented, and finally generating subgraphs by using the Graph Database added the complemented edges. For this study, we generated questions in the field of history, and since history requires temporal and spatial panoramic knowledge, words related to these aspects were focused on and complemented the relationships between them. As a result, 2,746 relationships were complemented by the proposed method in the subgraphs, and the subgraphs contained more words to learn (words found in textbooks that need to be learned) in a specific field compared to those generated using existing methods.