Memes and other text-embedded images are powerful tools for expressing opinions and identities, especially within marginalized socio-political movements. Detecting hate speech in this type of multimodal content is challenging because of the subtle ways text and visuals interact. In this paper, we describe our approach for Subtask A of the Shared Task on Multimodal Hate Detection in Marginalized Movement@CASE 2025, which focuses on classifying memes as either Hate or No Hate. We tested both unimodal and multimodal setups, using models like DistilBERT, HateBERT, Vision Transformer, and Swin Transformer. Our best system is the large multimodal model Qwen2.5-VL-7B-Instruct-bnb-4bit, fine-tuned with 4-bit quantization and instruction prompts. While we also tried late fusion with multiple transformers, Qwen performed better at capturing text-image interactions in memes. This LLM-based approach reached the highest F1-score of 0.8086 on the test set, ranking our team 5th overall in the task. These results show the value of late fusion and instruction-tuned LLMs for tackling complex hate speech in socio-political memes.

Identification of hate speech in images with text is a complicated task in the scope of online content moderation, especially when such talk penetrates into the spheres of humor and critical societal topics. This paper deals with Subtask A of the Shared Task on Multimodal Hate, Humor, and Stance Detection in Marginalized Movement@CASE2025. This task is binary classification over whether or not hate speech exists in image contents, and it advances as Hate versus No Hate. To meet this goal, we present a new multimodal architecture that blends the textual and visual features to reach effective classification. In the textual aspect, we have fine-tuned two state-of-the-art transformer models, which are RoBERTa and HateBERT, to extract linguistic clues of hate speech. The image encoder contains both the EfficientNetB7 and a Vision Transformer (ViT) model, which were found to work well in retrieving image-related details. The predictions made by each modality are then merged through an ensemble mechanism, with the last estimate being a weighted average of the text- and image-based scores. The resulting model produces a desirable F1- score metric of 0.7868, which is ranked 10 among the total number of systems, thus becoming a clear indicator of the success of multimodal combination in addressing the complex issue of self-identifying the hate speech in text-embedded images.

The rapid spread of misinformation on social media platforms necessitates robust detection mechanisms, particularly for languages with limited computational resources. This paper presents our system for the DravidianLangTech 2025 shared task on Fake News Detection in Malayalam YouTube comments, addressing both binary and multiclass classification challenges. We propose a Multi-Pooling Feature Fusion (MPFF) architecture that leverages [CLS] + Mean + Max pooling strategy with transformer models. Our system demonstrates strong performance across both tasks, achieving a macro-averaged F1 score of 0.874, ranking 6th in binary classification, and 0.628, securing 1st position in multiclass classification. Experimental results show that our MPFF approach with XLM-RoBERTa significantly outperforms traditional machine learning and deep learning baselines, particularly excelling in the more challenging multiclass scenario. These findings highlight the effectiveness of our methodology in capturing nuanced linguistic features for fake news detection in Malayalam, contributing to the advancement of automated verification systems for Dravidian languages.

Political discourse on social media platforms significantly influences public opinion, necessitating accurate sentiment analysis for understanding societal perspectives. This paper presents a system developed for the shared task of Political Multiclass Sentiment Analysis in Tamil tweets. The task aims to classify tweets into seven distinct sentiment categories: Substantiated, Sarcastic, Opinionated, Positive, Negative, Neutral, and None of the above. We propose a Multi-View Feature Fusion (MVFF) architecture that leverages XLM-R with a CLS-Attention-Mean mechanism for sentiment classification. Our experimental results demonstrate the effectiveness of our approach, achieving a macro-average F1-score of 0.37 on the test set and securing the 2nd position in the shared task. Through comprehensive error analysis, we identify specific classification challenges and demonstrate how our model effectively navigates the linguistic complexities of Tamil political discourse while maintaining robust classification performance across multiple sentiment categories.

Detecting misogyny in multimodal content remains a notable challenge, particularly in culturally conservative and low-resource contexts like Bangladesh. While existing research has explored hate speech and general meme classification, the nuanced identification of misogyny in Bangla memes, rich in metaphor, humor, and visual-textual interplay, remains severely underexplored. To address this gap, we introduce BanMiMe, the first comprehensive Bangla misogynistic meme dataset comprising 2,000 culturally grounded samples where each meme includes misogyny labels, humor categories, metaphor localization, and detailed human-written explanations. We benchmark the various performance of open and closed-source vision-language models (VLMs) under zero-shot and prompt-based settings and evaluate their capacity for both classification and explanation generation. Furthermore, we systematically explore multiple fine-tuning strategies, including standard, data-augmented, and Chain-of-Thought (CoT) supervision. Our results demonstrate that CoT-based fine-tuning consistently enhances model performance, both in terms of accuracy and in generating meaningful explanations. We envision BanMiMe as a foundational resource for advancing explainable multimodal moderation systems in low-resource and culturally sensitive settings.