Dynamic Knowledge Graph with Descriptions for Reasoning

While the integration of knowledge graphs and language models has shown promise in question answering, existing methods face limitations: (1) simplistic semantic representations of entity relations struggle to capture implicit contextual information; (2) static graph structures often introduce redundant nodes, degrading reasoning accuracy. We propose DynamiKG-Desc (Dynamic Knowledge Graph with Descriptions for Reasoning), which dynamically injects entity descriptions from Wiktionary to build context-aware knowledge representations and employs an attention-based node filter to dynamically select relevant nodes, reducing noise. Experiments on the CommonsenseQA benchmark demonstrate that DynamiKG-Desc significantly outperforms baselines, validating the effectiveness of its dynamic semantic enhancement mechanism.