21th AIAI 2025, 26 - 29 June 2025, Limassol, Cyprus

Bi-Modal Image-Text Fusion for Sentiment Analysis

NOUISSER AICHA, KHEDIRI NOUHA, kherallah monji, charfi faiza

Abstract:

  Sentiment analysis of memes is crucial in domains such as finance and politics, but the focus has been mainly on English. This study presents our test dataset MemoSen, and a proposed bimodal system dedicated to memes, integrating both text and image, with three sentiment labels: positive, negative, and neutral. A detailed annotation manual is provided to facilitate the development of new resources in this area, thus promoting the extension of sentiment analysis to various languages. Our first model, which integrates a Convolutional Neural Network (CNN), branches to process both images and text using a concatenation merge, and achieves a validation performance of an Area Under the Curve (AUC) about 100%. The second model, based on VGG16, achieved an AUC of 99.56%, while the third model, based on ResNet50, achieved an AUC of 98.69%.  

*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.