go back

Memory Net: Generalizable Common-Sense Reasoning over Real-World Actions and Objects

Julian Eggert, Jörg Deigmöller, Pavel Smirnov, Johane Takeuchi, Andreas Richter, "Memory Net: Generalizable Common-Sense Reasoning over Real-World Actions and Objects", International Conference on Knowledge Engineering and Ontology Development, 2023.


Abstract. We address the problem of situated reasoning of artificial agents (AA) in human-like environments. In particular, we want the AAs to reason about so-called action patterns in a real-world human environment: E.g., which tools can be used for a certain action, which actions can be performed with certain tools and objects, and so on. This should occur in a situated way, i.e., always referring to concrete instances in a real-world environment. For this purpose, we develop a representational concept based on knowledge graphs. We consistently populate a large- scale knowledge graph (MemNet) with densely interrelated everyday concepts, common-sense action patterns, and environment information. We then propose an inference mechanism that uses semantic proximity by approximate subgraph matching, which allows broad semantic generalization. As a proof-of-concept, the system is compared with human data and a state-of-the-art Natural Language Understanding (NLU) based machine learning approach, responding tool questions in a kitchen environment setting. The results show that the Knowledge Graph approach is able to generalize well for this type of tasks, with considerable benefits in terms of applicability in variable, incremental and interactive settings.

Download Bibtex file Download PDF