- 學位論文
於有雜訊的知識圖譜中進行知識圖譜推論
Learning logics rules and reasoning on noisy knowledge graphs
摘要
知識圖譜經常被用於不同的領域，例如商業中的問答系統和推薦系統。除此之外，人們還希望可以進一步去了解這些系統給出的答案背後的原因。因此，知識圖譜推論的相關研究就越來越受歡迎。 然而，大多數知識圖譜應用都假設知識圖譜是完全無雜訊的，而忽略現實世界中大多數的知識圖譜並非如此。事實上，知識圖譜的建構大多是從研究文檔或文章中提取出來，因此極容易發生萃取錯誤的問題；此外，知識發展的迅速也使得文獻當中的假設會被推翻，然而被推翻的關係依舊會留在知識圖譜當中，造成知識圖譜含有不正確的資訊，進而影響到知識圖譜推論的結果。在這項研究中，我們嘗試在知識圖推理模型上融合兩種加權的方法，以使模型在有雜訊的知識圖數據集上仍舊可做出良好的預測結果。 在本研究中，我們針對預測過程中的推論路徑進行合理性及邏輯性的評估。在實驗上，以預測而言，我們的模型在有雜訊的生物醫學知識圖譜上表現優於原本的模型。此外，為進一步驗證模型的能力，我們在生物醫學數據集上加入模擬雜訊，實驗結果顯示，我們提出的方法的效能除了優於原本的模型，同時也優於純類神經模型 RotatE。另一方面，針對預測過程中的推論路徑，評估結果顯示模型能夠為預測結果提供高度合理且符合邏輯的解釋。
並列摘要
Knowledge graph has frequently been utilized in different fields and domain, such as question-answering and recommendation systems in the business context. In addition, for a given question, people expect to know the reasons behind an answer derived from a knowledge graph for a specific question. Hence, reasoning over a knowledge graph has gained great research attention in recent years. However, most applications of knowledge graphs assume that the knowledge graph is completely noise-free, ignoring the fact that most knowledge graphs in the real world are extracted from research documents or articles and thus often contain noises (i.e., incorrect triplets). Besides, the evolution of knowledge could quickly overthrow theories, leaving knowledge graphs to contain contradicted and incorrect triplets and, in turn, hinder the effectiveness of reasoning over these knowledge graphs. Accordingly, in this research, we attempt to apply two weighting techniques to a knowledge graph reasoning model to make the model perform well on noisy knowledge graph datasets. In this research, we evaluate our proposed knowledge graph reasoning model using entity prediction as the evaluation task. The experimental results suggest that our proposed model outperforms the salient benchmark model on a noisy biomedical knowledge graph and also outperforms RotatE on simulated biomedical datasets. Moreover, our model is capable of providing highly interpretable logic rules for prediction results.