Rel4KC: A Reinforcement Learning Agent for Knowledge Graph Completion and Validation
Reinforcement Learning (RL) has been recently adopted to train agents for knowledge graph completion tasks on structured database. However, new fact triples extracted through non- community contribution added to the database for completeness could be invalid due to noise in the input data and limitation of relationship discovery algorithm itself. In this study, we propose Rel4KC, a RL agent that learns from massive structured data and then performs completeness and correctness checking on the triple facts extracted from free text through neural relation extraction. Reward shaping based on embeddings of entities and relations are used to enhance RL agent’s performance. The numerical experiments for a real-world problem demonstrate that the proposed approach yields promising results and has the potential of being adopted in knowledge graph generation and validation flow to ensure the trustworthiness of triple facts being populated into the knowledge base. Furthermore, a Japanese subset of the knowledge database is used to validate the multilingual extensibility of prototyped agent.
READ FULL TEXT