Reinforcement Learning for Dynamic RDF Schema Evolution in NoSQL Databases

Abstract

Reinforcement learning offers a promising foundation for adaptive schema extraction from heterogeneous NoSQL data. This work presents a dynamic framework for RDF schema evolution, where schema construction is modeled as a sequential decision-making process. Semantic representations of extracted (Subject, Predicate, Object) triplets are computed using Sentence-BERT embeddings, capturing contextual and lexical similarities across documents. These embeddings inform the decisions of a reinforcement learning agent, which incrementally updates the schema by selecting among actions—Add, Merge, Modify, or Ignore—based on long-term optimization of semantic coherence, structural compactness, and constraint compliance. A reward function guides policy learning by incorporating constraint violations and redundancy metrics. The framework also supports the continuous inference and refinement of cardinality and datatype constraints, enabling the schema to adapt over time to evolving data structures in schema-less NoSQL environments.