Deep Learning Multi-head Attention Based Fake News Detection for the Amharic Language

Abstract

Due to the increasing of social media usage, it has become necessary to combat the spread of false information and decrease the reliance on information retrieval from such sources. Social platforms are under constant pressure to come up with an efficient method to solve this problem because users' interaction with fake and unreliable news leads to its spread at an individual level. This spreading of misinformation adversely affects the perception of important activity, and as such, it needs to be dealt with using a modern approach. This research presents a deep learning based fake news detection using a multi-head attention for Amharic language. This research proposes multi head attention-based BI-LSTM, ensemble of CNN-LSTM, and LSTM which are state-of-the-art and NLP approaches. The goal of this research is to analyze Amharic text and classify news taken from social media as fake or real. To evaluate the news classification system, this research collects about 2400 news from different social and Main stream Media. This research also tries to investigate the effect of removing Amharic language stop words during preprocessing stage on the Amharic language fake news detection process. The proposed multi-head attention model outperforms other deep learning models such as BI LSTM, CNN, LSTM, and CNN-LSTM by 7%, 7%, 16%, and 7% respectively. The BI-LSTM, CNN, CNN-LSTM produces almost equal classification accuracy. Results shows that applying stop words drop the accuracy of the fake news detection by 2.83% while examine on propose attention-based.