Dynamic Knowledge Graph Augmentation Enhances Factual Accuracy in Retrieval Based Generation

Large language models (LLMs) demonstrate exceptional fluency in natural language generation but remain susceptible to producing factually incorrect outputs due to static parametric knowledge frozen at training time. Retrieval-augmented generation (RAG) partially mitigates this limitation by conditioning generation on externally retrieved evidence, yet conventional RAG systems depend on unstructured, flat document corpora that fail to represent the relational and temporal dynamics characterizing real-world knowledge. This paper proposes Dynamic Knowledge Graph Augmentation (DKGA), a framework that integrates continuously updated knowledge graphs (KGs) with retrieval-based generation pipelines to substantially improve factual accuracy.

DKGA employs a graph neural network (GNN) encoder for subgraph-conditioned entity representation learning, a temporal update module for incremental knowledge refresh, and a cross-modal relevance-aware fusion mechanism that jointly conditions the generator on structured KG evidence and unstructured text passages. Experiments on the WebQA and TriviaQA benchmarks demonstrate that DKGA achieves an 11.4% improvement in exact match factual accuracy and an 18.2% reduction in hallucination rate over strong RAG baselines. These results provide compelling evidence that dynamic, structured knowledge representations are a critical and underutilized resource for knowledge-intensive language generation.