用於知識圖譜推理的語義導向規則學習

Abstract

隨著知識圖譜廣泛應用於問答系統、醫療推論與推薦等場域，具可解釋性的推理方法需求日益提升。相較於嵌入式模型的黑盒特性，部分神經推理方法能結合符號規則，提供清晰的邏輯推理，尤其在生醫等知識密集領域更具可信度。然而，這類方法在處理大型異質圖譜時常缺乏語義感知能力，導致所學規則語意品質不穩。 本研究提出 SemRNNLogic方法，在原始 RNNLogic 架構上引入語義輔助任務，結合多任務學習設計，利用生醫圖譜中的實體語義型別資訊，提升模型對規則生成的語義理解與推理能力。具體來說，我們於規則生成的每一步預測過程中，設計兩項語義輔助任務：一為預測下一個關係所屬的語義群組，一為預測下一個實體在推理路徑上的語義型別分布，以強化語意導向的序列生成。 實驗結果顯示，語義導向的多任務模型在 SemMedDB 資料集上整體提升連結預測效能，特別在 MRR、Hit@1 與 Hit@3 等前段排序指標上優於原始模型，顯示語義輔助任務能有效增強規則學習的準確性。

As knowledge graphs are increasingly applied in areas such as question answering, medical reasoning, and recommendation systems, the demand for interpretable reasoning methods continues to grow. Compared to the black-box nature of embedding-based models, some neural-symbolic reasoning approaches incorporate symbolic rules to provide clearer logical inference, which is particularly valuable in knowledge-intensive domains like biomedicine. However, these methods often lack semantic awareness when dealing with large knowledge graphs, leading to unstable semantic quality in the generated rules. In this work, we propose SemRNNLogic, an enhanced version of the original RNNLogic architecture that incorporates semantic auxiliary tasks through a multitask learning framework. By leveraging semantic type information of entities in biomedical KGs, our model aims to improve the semantic understanding and reasoning capabilities during rule generation. Specifically, we design two auxiliary tasks at each step of rule generation: one classifies the next relation into a semantic cluster, and the other one predicts the semantic type distribution of the next entity in the grounding path, guiding the sequence generation process with semantic cues. Experimental results show that our semantic-guided multitask model improves overall link prediction performance on the SemMedDB dataset. In particular, it achieves better scores than the original model on top-ranked metrics such as MRR, Hit@1, and Hit@3, demonstrating that semantic auxiliary tasks effectively enhance rule learning accuracy.