|
Explainable Text Classification Model for Mumtahina Ahmed1, Mohammad Shahadat Hossain2, Raihan
Ul Islam3, and Karl Andersson3+
1Department of Computer Science and Engineering Port City
International University 2Department of Computer Science and Engineering University of
Chittagong 3Department of Computer Science, Electrical and Space Engineering Luleå University
of Technology, Skellefteå,
Sweden Abstract Artificial intelligence has achieved
notable advances across many applications, and the field is recently
concerned with developing novel methods to explain machine learning models.
Deep neural networks deliver the best performance accuracy in different
domains, such as text categorization, image classification, and speech
recognition. Since the neural network models are black-box types, they lack
transparency and explainability in predicting results. During the COVID-19
pandemic, Fake News Detection is a challenging research problem as it
endangers the lives of many online users by providing misinformation.
Therefore, the transparency and explainability of COVID-19 fake news
classification are necessary for building the trustworthiness of model
prediction. We proposed an integrated LIME-BiLSTM model where BiLSTM assures
classification accuracy, and LIME ensures transparency and explainability. In
this integrated model, since LIME behaves similarly to the original model and
explains the prediction, the proposed model becomes comprehensible. The
performance of this model in terms of explainability is measured by using
Kendall’s tau correlation coefficient. We also employ several machine
learning models and provide a comparison of their performances. Therefore, we
analyzed and compared the computation overhead of our proposed model with the
other methods because the model takes the integrated strategy. The
comprehensible model for fake news detection outperforms other models with
94.25% accuracy, and Kendall’s tau correlation value of 0.35. Keywords: fake news, COVID-19,
Explainable AI, LIME, BiLSTM +: Corresponding author: Karl Andersson Journal of Internet Services and
Information Security (JISIS), 12(2): 51-69, May 2022 |