Towards Explainable Sarcasm Detection in Low-Resource Language using Bi-LSTM

Sarcasm means using humor to mock someone or something. The task of understanding sarcasm is very challenging because it carries double meaning and can often lead to confusion and misunderstanding. Hence the need for differentiating sarcasm from normal text is necessary. As the number of users in Social Networks Sites (SNS) is increasing day by day, the volume of data online is also going beyond our control. So automatic and accurate detection of sarcastic comments from this huge data is extremely necessary. While sarcasm detection has been researched for many years in English and other languages, it still remains an underresearched topic in low resource languages like Bengali. We have combined two benchmark datasets BanglaSarc and BenSarc into one to produce a larger dataset. Then we implemented Deep Learning models on this custom dataset to observe their performances. The experimental results show the superiority of Bidirectional Long-Short Term Memory (Bi-LSTM) model without pre-trained word embedding over all other models with an accuracy of 72.10%, precision 67.62%, recall 65.44% and fl-score 66.52% respectively. The use of Local Interpretable Model-Agnostic Explanations (LIME), an explainable AI technique, has made the results more interpretable.