A Method to Constract a Masked Knowlege Graph Model using Transformer for Knowledge Graph Reasoning
Ryoya Kaneda, Makoto Okada, Naoki Mori · 2023
Most of the previous methods using machine learning for this challenge generate a new knowledge graph from the original one, and some information is lost in the process of creating a new knowledge graph. Therefore, we proposed a new model to estimate the criminal without changing the original knowledge graph. The proposed model uses a Transformer and allows the estimation of unknown criminals in nonexistent scenes by learning similar to Masked Language Modeling in BERT. This model, which uses the original knowledge graph, is expected to infer information about the crime scene at the same time as predicting the criminal. We confirmed by experiments that the model had gained the ability to estimate the hidden story parts by considering the surrounding stories.