Fake News Detection in Indian Languages: A Case Study with Hindi Using CNN-LSTM

The increasing spread of false information also known as fake news has had a negative impact on the society and the political system hence the need for detection tools. This research work presents a hybrid CNN-LSTM deep learning model for detecting fake news in Hindi, a language that lacks adequate dataset and resources. A new dataset of 6,724 Hindi news articles (2,704 fake and 4,020 real) was collected from the trusted sources which are members of International Fact Checking Network (IFCN). The model uses FastText pretrained embeddings, a Conv1D layer for local feature extraction and LSTM units for sequential feature extraction, and is able to achieve 97% accuracy on the proposed dataset and an F1 score of 89% on CONSTRAINT2021 dataset. This paper also presents a new dataset for future research and the first work done towards developing a system for detecting fake news in Hindi language. In the future, the work will be continued by trying to apply this approach to other sparse Indian languages and by using transformer-based models to improve results.