Izvorni znanstveni članak
https://doi.org/10.32985/ijeces.17.4.3
Efficient Approach of Sarcasm News Headlines Segregation using LSTM and LDA Topics Analysis in Recurrent Neural Network
Anantha Babu S
; Department of Computer Science and Engineering, GITAM School of Technology, GITAM University, Bangalore, India
A. Shobanadevi
; Department of Data Science and Business Systems, School of Computing, SRM Institute of Science and Technology, Kattangulathur, Chennai
A. Rajasekaran
; Department of Artificial Intelligence and Machine Learning, Rajalakshmi Engineering College, Thandalam, Chennai, India
Umanesan R
; Department of Computer Science and Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Chennai, India
Samsudeenshaffi S
orcid.org/0000-0001-5369-7946
; Department of Computer Science and Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Chennai, India
Jagadeesh S
orcid.org/0000-0001-8907-4433
; Department of Computer Science and Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Chennai, India
*
* Dopisni autor.
Sažetak
The increasing spread of misinformation on social media highlights the importance of sarcasm detection, as sarcastic expressions often obscure the real intent of a message and hinder accurate classification. This work combines Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) models, and Latent Dirichlet Allocation (LDA) to develop a robust framework for detecting sarcasm in news headlines. The approach applies text preprocessing techniques such as tokenisation, stop-word removal, lemmatisation, and stemming, followed by topic modelling and evaluation using Jensen–Shannon divergence. Experimental analysis shows that the proposed hybrid CNN–RNN (LSTM) model, strengthened with GRU blocks, regularisation (Lasso and Ridge), dropout, and batch normalisation, achieves 99% accuracy in sarcasm prediction. The proposed architecture delivers a significant improvement compared to traditional machine learning baselines like logistic regression and SVMs, which typically achieve 70–80% accuracy, as well as prior deep learning models such as standalone CNNs or LSTMs that report accuracy in the 85–99% range. In addition, the integration of topic modelling produces more coherent clusters and better resilience to class imbalances. These findings demonstrate that combining topic modelling with deep neural architectures provides a highly effective strategy for sarcasm detection and can support more reliable misinformation analysis on social platforms.
Ključne riječi
Natural Language Processing; Deep Learning; LSTM. LDA; CNN-RNN;
Hrčak ID:
345970
URI
Datum izdavanja:
1.4.2026.
Posjeta: 0 *