CUET_NetworkSociety@DravidianLangTech 2025: A Multimodal Framework to Detect Misogyny Meme in Dravidian Languages - Abstract & Introduction

Sep 20, 2026

Memes are widely used to communicate on social networks, and some convey misogynistic, harmful messages. This paper proposes a multi-modal deep learning framework to detect misogynistic memes in Tamil and Malayalam, addressing the difficulty of low-resource language morphology. The model fuses fine-tuned ResNet18 visual features with ai4bharat/indic-bert text representations, then evaluates on a test dataset. Results report macro F1 of 76.32% for Tamil and 80.35% for Malayalam, and competitive ranking outcomes.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages7

Reading Time18 min

AccessFree to download

Abstract

1 Introduction

2 Related Work

CUET_NetworkSociety@DravidianLangTech 2025: A Multimodal Framework to Detect Misogyny Meme in Dravidian Languages MD Musa Kalimullah Ratul* , Sabik Aftahee* , Tofayel Ahmmed Babu* Jawad Hossain and Mohammed Moshiul Hoque Department of Computer Science and Engineering Chittagong University of Engineering and Technology {u1904071, u1904024, u1904005, [u1704039}@student.cuet.ac.bd](u1704039}@student.cuet.ac.bd)[ ](u1704039}@student.cuet.ac.bd)[moshiul_240@cuet.ac.bd](moshiul_240@cuet.ac.bd) Abstract Memes are widely used to communicate on social networks. Some memes can be misogynistic, conveying harmful messages to individuals. Detecting misogynistic memes has become a broader challenge, particularly for low-resource languages like Tamil and Malayalam, due to their linguistic morphology. This paper presents a multi-modal deep learning approach for detecting misogynistic memes in Tamil and Malayalam. The proposed model integrates fine-tuned ResNet18 for extracting visual features and ai4bharat/indic-bert for analyzing textual content. The fusion model was then applied to make predictions on the test dataset. The model achieved a macro F1 score of 76.32% for Tamil and 80.35% for Malayalam. Our technique helped secure 7th and 12th positions for Tamil and Malayalam, respectively. 1 Introduction In the past few years, the popularity of the social media platform has gained a huge response from individuals, where a multi-modal phenomenon called meme has been introduced to us. The meme is generally an image with some contextual texts of that image. Apart from humorous contents, memes also can carry harmful messages, such as misogyny, which is the hatred of, contempt for, or prejudice against women or girls. The satirical nature of such content makes identifying misogynistic memes particularly challenging, as they often employ nuanced, context-dependent signals that can evade straightforward detection (Rizzi et al., 2023) . To illustrate this distinction, Figure 1 compares a misogynistic meme and a non-misogynistic meme, highlighting the subtle yet significant differences in their messaging. Misogynistic memes pose a threat to society that basically try to normalize hatred against women and gender bias. Existing research has explored multimodal approaches to identifying misogyny contents in memes with both textual and visual features to achieve higher accuracy. For instance, Rizzi et al. (Rizzi et al., 2023) investigated 4 unimodal and 3 multimodal approaches to determine which source of information contributes more *Authors contributed equally to this work. A misogynistic meme A non-misogynistic meme Figure 1: Examples of memes: (a) Misogynistic and (b) Non-misogynistic. to the detection of misogynous memes. Similarly, Singh et al. (Singh et al., 2024) investigated misogyny detection in Hindi-English code-mixed memes, showcasing the complexities of handling low-resource languages and multimodal content. The shared task on multitask meme classification further emphasized the importance of distinguishing misogynistic and trolling behaviors, providing valuable datasets and benchmarks for advancing the field (Chakravarthi et al., 2024) . Even with this much advancement, research on misogynistic memes in low-resource languages such as Tamil and Malayalam remains very much challenging. Our paper introduces a multi-modal deep learning framework for detecting misogynistic memes in Tamil and Malayalam. The key contributions of this work include: • Proposed a multimodal framework designed to detect misogynistic memes while capturing the linguistic and cultural nuances of Tamil and Malayalam. • Investigated the performance of several ML, DL, and transformer-based models for misogynistic meme detection, highlighting challenges in lowresource languages through quantitative and qualitative error analysis. 2 Related Work Multimodal approaches, which leverage the interplay between text and images, have shown promis

FAQs

What problem does the paper address?

The paper addresses detecting misogynistic memes on social networks, focusing on low-resource Dravidian languages such as Tamil and Malayalam.

How does the proposed approach combine image and text information?

It extracts visual features using a fine-tuned ResNet18 and analyzes textual content using ai4bharat/indic-bert, then fuses both for prediction.

What performance results does the model achieve?

The model reports macro F1 scores of 76.32% for Tamil and 80.35% for Malayalam on the test dataset.

Sep 20, 2026

Uploaded by supergirl