利用參考引導和領域不變的後置軟提示進行跨領域假新聞檢測

Abstract

由於假新聞的多樣性和不斷變化的主題，在早期偵測假新聞面臨重大挑戰。現有方法依賴於難以獲取的特徵或排除特定領域的特徵，這可能會限制其性能。在本文中，我們介紹了一種新穎的方法，利用相關新聞作為參考來早期識別假新聞。我們的方法利用軟提示調整生成兩種特徵：跨不同領域捕捉假新聞共同特徵的領域不變特徵，以及透過假新聞相關文章產生的參考特徵。接著，我們通過動態調整兩種特徵的比例來生成假新聞特徵，並以此來判別假新聞。我們的方法可以以零樣本的方式適應不同的主題或時期，而無需人工製作或難以獲取的特徵。為了評估我們方法的有效性，我們在包含中文和英文數據的兩個數據集上進行了實驗。結果表明，我們的方法在虛假新聞早期檢測方面超過了最先進的方法。

Detecting fake news in its early stage poses a significant challenge due to its diverse nature and ever-changing topics. Existing methods rely on difficult-to-acquire features or eliminate domain-specific features, which may limit their performance. In this paper, we introduce a novel method that utilizes related news as references to identify fake news at an early stage. Our approach leverages soft prompt tuning to generate two features: domain-invariant features that capture common characteristics of fake news across various domains and reference features that capture the specific context of each fake news instance with external reference articles. Next, we generate fake news features by dynamically adjusting proportions of the two types of features and use them to determine fake news. Our method can adapt to different topics or periods in a zero-shot manner without needing hand-crafted or hard-to-get features. To evaluate the effectiveness of our approach, we conduct experiments on two datasets comprising Chinese and English language data. The results demonstrate that our method surpasses state-of-the-art techniques in fake news early detection.