社群媒體包含多樣且未經過濾的訊息，為了避免傳播未經證實的主張，這使得自動事實查核被迫切需要。自動事實查核旨在考慮特定證據的狀況下，決定主張之真實性。現有方法通常遵循兩種方法，包括基於蘊涵的三元分類方法或典型多類別分類等設置，這兩種方法對於真實情況都有其不足之處。在本文中，我們提出了一種新穎的模型，據我們所知，這是第一個利用異常偵測原理進行事實查核的模型。我們為異常實例設計額外的損失函數，並且成功地使特定類別的學習更加專注。實驗結果顯示我們提出的模型可以表現更佳或具有競爭力的結果，這證明了在兩個基準數據集，FEVEROUS 和 LIAR 資料集上的有效性。;Social media contains diverse and unfiltered information, which makes automatic fact verification in urgent need against the dissemination of unproven claims. Fact Verification aims to assign an authenticity verdict to a claim in particular evidence. Existing approaches usually follow the setting of entailment-based ternary or typical multi-class classification, which are not feasible for realistic scenarios. In this paper, we propose a novel model, which is, to our best knowledge, the first to exploit out-of-distribution detection for fact checking. By incorporating an extra keep-away loss for out-of-distribution instances, we successfully engage the learning with respect to specific categories. Experimental results show our model performs better or competitive results, which demonstrates the effectiveness on two benchmark datasets, the FEVEROUS and LIAR dataset.
