植基於深度學習假新聞人工智慧偵測：台灣真實資料實作

Deep Learning Based Fake News AI Detection：Evidence From Taiwan News Report

近年來透過網路媒介，使得假新聞得以快速流竄。而許多國家都受到假新聞嚴重影響，使得假新聞偵測變成一個重要課題。本研究蒐集兩個台灣闢謠網站資料，「新聞小幫手」以及「真的假的」。並運用深度學習的三種方法Gated Recurrent Unit（GRU）、Long Short Term Memory（LSTM）以及Bidirectional Long Short Term Memory（BLSTM），進行假新聞偵測。實驗結果發現，深度學習可以運用在台灣假新聞偵測當中，且BLSTM的方法最適合用於假新聞偵測。本研究實驗降低假新聞比例，模擬25％以及5％假新聞比例，讓研究樣本可以更趨近於真實情況。最後本研究使用交叉資料集測試，了解實務與學術上的差距。

In recent years, because the Internet acts as a medium, fake news can be quickly spread. Many countries have been seriously affected by fake news. making fake news detection become an important issue. This study collects two Taiwan refute rumors sites. And use the three methods of deep learning for fake news detection., Gated Recurrent Unit (GRU), Long Short Term Memory (LSTM), and Bidirectional Long Short Term Memory (BLSTM). The experimental results show that deep learning can be used in Taiwan's fake news detection, and the BLSTM method works best. Research experiments reduce the proportion of fake news, simulating 25% and 5% fake news ratios. Let the research sample be closer to the real situation. Finally, this study used a cross-data set test to understand the gap between practice and theory.