Performance Analysis of Hybrid Deep Learning Models in Sarcasm Classification
Md Fatin Faiaz Isty, S. M. Mahedy Hasan, Md Shamiul Islam Shopnil · 2022
Sarcasm as an apparatus for exhibiting dissatisfaction, vexation, cynicism, shrewdness, mockery of ignorance, and foolishness has become ubiquitous in textual and electronic media. A sarcasm detection model, an automated method for classifying sarcastic data from multiple sources, is needed as a result of this effect of sarcasm on human sentiment perception. But the hardships associated with textual data processing apply to this task as well. For that reason, our work can be considered a three-phase classification method, where, the first phase is concerned with the assortment of text data which consist of Noisy and Well Structured text data. In the second phase, a method was devised for the conversion of textual data into numerical entities. The third phase is concerned with structuring our deep learning method for the precise classification of sarcastic data. We used FastText in this work to solve contextual problem due to information removal from dirty text data. FastText was employed with a standard CNN model, which was very quick but had generalization issues with 90.17% accuracy. Later, FastText was combined with an LSTM model, which improved generalization with 92.16% accuracy but costly in time consumption. In this study, we proposed a FastText model with CNN and LSTM that makes use of CNN’s advantages in terms of efficiency and LSTM’s advantages in terms of generalization with an accuracy of 91.29%.