- 學位論文
偵測錯誤健康新聞的階層式圖注意力網路
Hierarchical Graph Attention Network for Fake Health News Detection
摘要
隨著社群平台的興起，大量錯誤的醫療健康新聞流傳於網際網路上，當人們採取健康假訊息建議的偏方後，他們的生命可能會受到威脅。為了避免假新聞造成的負面影響，許多偵測的方法已被提出，例如，自然語言處理技術（NLP）能夠根據新聞的文字來判斷其真實性，然而由於當今人們時常從社群媒體接收新聞資訊，用戶的背景以及其對新聞的參與模式或許有助於假新聞的偵測，因此，研究學者引入圖神經網路（GNN）到此任務上。通常在一個社群網路中，每個節點對他相鄰的節點有不同的影響力，每種關係也有獨特的意義，有鑑於此，我們提出一個新穎、以階層式注意力機制為基礎的圖學習框架，以捕抓重要的節點和交互作用。另外，因為圖神經網路在多層堆疊時表現不佳，我們設計了兩階段的訓練策略，以縮短傳遞用戶交友圈之訊息到新聞節點的路徑。在辨別健康假新聞的任務上，實驗結果顯示我們的模型優於現有的方法，並且基於注意力機制的圖神經網路能受益於兩階段的訓練。
並列摘要
With the rise of social media, massive fake health news is flooding over the Internet. When people take the treatments suggested by health misinformation, their lives may be at risk. To prevent the negative impacts caused by fake news, numerous detection methods have been proposed. For example, Natural Language Processing (NLP) techniques have been utilized to debunk fake news based on the content in the story. However, since nowadays people often receive information from social media, the background of platform users and their engagement with news may be useful for news verification. Researchers thus introduce Graph Neural Networks (GNNs) to this task. Generally in a social network, each node has different influences on its adjacent nodes and each relationship represents a unique meaning. As a result, we propose a novel graph learning framework based on hierarchical attention to capture the important nodes and interactions. Furthermore, GNNs often perform poorly when stacking with multiple layers, so we design the two-stage training strategy to shorten the paths of delivering user network information to news nodes. The experimental results show that our model outperforms the existing methods in the task of health misinformation detection and attention-based GNNs can be benefited from the two-stage training.