Title: Optimal pre-trained deep ensemble of classification model for multimodal sarcasm detection

Authors: Dnyaneshwar Madhukar Bavkar; Ramgopal Kashyap; Vaishali Khairnar

Addresses: Department of Computer Science & Engineering, Amity University, Raipur, Chhattisgarh, India ' Department of Computer Science & Engineering, Amity University, Raipur, Chhattisgarh, India ' Department of Information Technology, Terna Engineering College, Nerul, Navi Mumbai, India

Abstract: The practice of utilising words or sentences that have a meaning other than their literal meaning is known as verbal irony or sarcasm. Making a machine recognise sarcasm is not an easy task because it can take humans a while to comprehend it. Deep Learning (DL) is becoming increasingly necessary for operations involving detection and classification. The four essential steps in this method are pre-processing, feature extraction, improved modality level fusion and ensemble classification technique. The very next stage is feature extraction, wherein n-gram, cosine similarity and improved TF-IDF (ITF-IDF) features are extracted from the text. Through improved modality level feature fusion, all the input modality attributes that have been extracted are placed through fusion to produce the fused feature set. An ensemble classification model is proposed that uses Deep Maxout, DBN, CNN and Bi-LSTM. Atom Search Assisted Bald Eagle Optimisation (ASBEO) trains the Bi-LSTM by tuning the optimum weights.

Keywords: sarcasm detection; ensemble model; deep learning; optimisation; tokenisation.

DOI: 10.1504/IJWMC.2026.150855

International Journal of Wireless and Mobile Computing, 2026 Vol.30 No.1, pp.10 - 31

Received: 12 Jan 2023
Accepted: 01 Mar 2024
Published online: 24 Dec 2025 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article