Contextualized Scene Knowledge Graphs for XAI Benchmarking.

In order to utilize artificial intelligence (AI) safely and securely in society, explainable artificial intelligence (XAI) technology, which has the property of being able to explain the reasons why a system has reached a conclusion, is necessary. Therefore, although machine learning approaches are currently the mainstream of AI, AI technology that combines inductive machine learning and deductive knowledge utilization is expected to become necessary in the future. Currently, however, there is no dataset to evaluate both approaches properly. In this study, we constructed and refined large-scale scene graphs and event-centered knowledge graphs, and have released them as open data. While most knowledge graphs contain only simple relationships, the constructed knowledge graphs are characterized by the fact that they contain more complex relationships that reflect the real world, such as temporal, causal, and probabilistic relationships. In addition, we developed refinement methods for the actual use of the constructed knowledge graphs for inference and machine learning. We held four technical competitions in Japan for AI technologies with various explanatory possibilities, gathered methods related to inference and estimation from a wide range of IT engineers, and classified the proposed technologies. An international version of the competition is planned for FY2022. In the future, we would like to design appropriate indices and conduct objective evaluations, classifications, and systematization for the development of AI technologies with explanatory properties, especially those that combine inductive machine learning (inference) and deductive knowledge utilization (reasoning).