The differentiable neural computer (DNC) is a neural network model with an addressable external memory that can solve algorithmic and question-answering tasks. Improved versions of the DNC have been proposed, including the robust and scalable DNC (rsDNC) and DNC-deallocation-masking-sharpness (DNC-DMS). However, integrating structured knowledge and calculations into these DNC models remains a challenging research question. In this study, we incorporate an architecture for knowledge and calculations into the DNC, rsDNC, and DNC-DMS to improve their abilities to generate correct answers for questions with multi-hop reasoning and provide calculations over structured knowledge. Our improved rsDNC model achieves the best performance for the mean top-1 accuracy, and our improved DNC-DMS model scores the highest for the top-10 accuracy in the GEO dataset. In addition, our improved rsDNC model outperforms other models in regards to the mean top-1 accuracy and mean top-10 accuracy in the augmented GEO dataset.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.