M.S. AAI Capstone Chronicles 2024

Detecting Fake News Using Natural Language Processing we utilized LIME for some explanations. Since the dataset is quite diverse, there was found to be certain language used in the majority of fake news articles that cued our model into predicting quite accurately. For example, one common theme in fake news that our team noticed was the lack of a professional tone, such as referring to Hillary Clinton as just “Hillary”. The word “Hillary” is flagged as fake, while “Clinton” is not because it is perceived to be more professional when referring to a person previously mentioned by their last name. Our team also noticed that words getting repeated frequently were also flagged as fake. In one example, shown below “FBI” is repeated 10 times in one paragraph.

Figure 9: The LIME explanation from our two-layer bidirectional LSTM when a fake news article is passed through. Since our project’s goal was to help users identify fake news and educate them on how to identify fake news in the future, performance is quite important. Achieving high accuracy and precision was important to our team. Accuracy is a measure of overall correctness in our model while precision is a measure of positive predictive values. It is more harmful to the user if a fake news article is labeled as correct (false positive) because misleading information can be spread throughout a community. There is often a tradeoff between recall and precision, which leads to

17

42