Sarcasm Detection: A Comparative Analysis of RoBERTa-CNN vs RoBERTa-RNN Architectures

Increasingly advanced technology and the creation of social media and the internet can become a forum for people to express things or opinions. However, comments or views from users sometimes contain sarcasm making it more difficult to understand. News headlines, sometimes contain sarcasm which makes readers confused about the content of the news. Therefore, in this research, a model was created for sarcasm detection. Many methods are used for sarcasm detection, but performance still needs to be improved. So this research aims to compare the performance of two text classification methods, Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN), in detecting sarcasm in English news headlines using RoBERTa text transformation.  RoBERTa produces a fixed-size vector of numbers 1x768. The research results show that CNN has better performance than RNN. CNN achieved the highest average accuracy of 0.891, precision of 0.878, recall of 0.874, and f1-score of 0.876, with a loss of 0.260 and a processing time of 508.1 milliseconds per epoch. On the contrary, RNN shows an accuracy of 0.711, precision of 0.692, recall of 0.620, f1-score 0.654, and loss of 0.564, with a longer processing time of 116500 milliseconds per epoch. The 10-fold cross-validation evaluation method ensures the model performs well and avoids overfitting. So it is recommended to use the combination of RoBERTa and CNN in other text classification applications that require high speed and accuracy. Further research is recommended to explore deeper CNN architectures or other architectural variations such as Transformer-based models for performance improvements.