A Federated Learning approach for Question and Answering on Knowledge Graphs

Abstract

Many organizations maintain Knowledge Graphs (KG) to store their data. To query this data, multiple Questions Answering (QA) systems have been proposed. In the scenario of multi-source Knowledge Graphs or a scenario where multiple organizations come together to develop a better and secure QA system to query their own data, a Federated Learning approach seems suitable to bridge this gap which enables the learning models to learn collaboratively across decentralized data sources, ensuring total data privacy and security. Hence, we propose our solution Fed-KGQA to address this problem. This research is limited to the domain of simple questions. Our research is subdivided into two main areas. First, we use a slightly improvised Federated Learning approach based on FedR which lets the clients train embeddings (of both entities and relations) locally using algorithms like TransE and then aggregate relation embeddings from all clients in a secure manner. Second, unlike traditional approaches, we employ KG Embedding (KGE) - based QA instead of SPARQL-based QA. Here we use the embeddings generated in the federated manner to represent entities and relationships. Given a question the system tries to infer the head, its embedding and the relational embedding from the query and then uses scoring function to deduce the tail entity, providing the answer to the user’s query. We test our approach and study its effectiveness on three subsets of FreeBase2M KG acting as clients with corresponding subsets of FreeBase SimpleQuestions as QA datasets. Our method shows an improvement in the accuracy as compared to plain QA on all clients while maintaining the security of clients and efficiency of the process.