Sarcasm Detection in Political Speeches Using Recurrent Neural Networks

This study investigated the effectiveness of recurrent neural networks (RNN) for the detection of sarcasm in the challenging domain of political speech. Given the inherent nuance of sarcasm, it can be difficult. This study compared three RNN architectures (SimpleRNN, LSTM, and GRU) and demonstrated that ensemble learning techniques (stacking and weighted averaging) further improved accuracy. Pre-trained word embeddings (GloVe) were used to capture semantic cues that signal sarcasm. These embeddings were incorporated by replacing words with their corresponding vector representations. The model was evaluated using standard metrics (accuracy, precision, recall, F1 score). The results showed that the ensembles outperformed individual RNNs, achieving a peak accuracy of 95% and an F1 score of 95% for both sarcastic and nosarcastic classes. Individual RNNs achieved an accuracy of 91%, highlighting the clear benefit of ensemble learning. This improvement suggests that ensembles effectively combine the strengths of different models, leading to more robust and generalisable sarcasm detection in political speech. Furthermore, this research paves the way for the application of similar techniques to sentiment analysis tasks in other complex domains.