Explainable Models for the Detection of Incidents of Fake News and Hate Speech

Authors:
  • VDVraj Desai
    Vraj Desai
  • AGAshray Gattani
    Ashray Gattani
  • HDHarshal Dalvi
    Harshal Dalvi

Abstract
The current existence of social media and its involvement in our society has granted people the freedom to develop new ways to interact and exchange information on a large scale leading to the spread of fake news and hate speech. Recently, there have been several AI-based analytical attempts developed that deal with the perception of fake news phenomena which uses the identification of typical fake news patterns, options, and hate speech detection. These AI-based models are unbiased in the sense that their features are arbitrarily chosen from the pool of accessible features. Using explainable artificial intelligence (XAI) we can counter these issues by providing more transparent and more clear explanations for the output given by traditional models. XAI techniques can help to identify a strong interconnection between the keywords and fake news detection patterns, essential for research toward achieving transparent user–machine interaction. In this chapter, we will be analyzing different XAI techniques available viz. LIME, SHAP, and eli5. The selected set of features are passed to global and local explainable models such as SHAP and LIME to generate probabilities for the features selected and plot the probabilities in a graphical format. This method can help machine learning (ML) practitioners to better understand the rationale behind model decisions and labeling a news article as fake/hate speech or as a genuine one. This can further help ML practitioners finetune the necessary parameters to improve model performance and address the trade-off between model accuracy and transparency. Experimental analysis of fake news and hate speech datasets using logistic regression determines that our proposed explainable model achieves remarkable classification of fake or real, hate or not hate.