This paper presents a novel approach to hate speech detection and target identification across Devanagari-script languages, with a focus on Hindi and Nepali. Leveraging an Attention BiLSTM-XLM-RoBERTa architecture, our model effectively captures language-specific features and sequential dependencies crucial for multilingual natural language understanding (NLU). In Task B (Hate Speech Detection), our model achieved a Macro F1 score of 0.7481, demonstrating its robustness in identifying hateful content across linguistic variations. For Task C (Target Identification), it reached a Macro F1 score of 0.6715, highlighting its ability to classify targets into “individual,” “organization,” and “community” with high accuracy. Our work addresses the gap in Devanagari-scripted multilingual hate speech analysis and sets a benchmark for future research in low-resource language contexts.

This study explores the challenges of natural language understanding (NLU) in multilingual contexts, focusing on Devanagari-scripted languages such as Nepali, Marathi, Sanskrit, Bhojpuri, and Hindi. Language identification within these languages is complex due to their structural and lexical similarities. We present a hybrid Attention BiLSTM-XLM-RoBERTa model, achieving a state-of-the-art F1 score of 0.9974 on the test set, despite limited resources. Our model effectively distinguishes between closely related Devanagari-scripted languages, providing a solid foundation for context-aware NLU systems that enhance language-specific processing and promote inclusive digital interactions across diverse linguistic communities.

This study addresses the challenge of fake news detection in code-mixed and transliterated text, focusing on a multilingual setting with significant linguistic variability. A novel approach is proposed, leveraging a fine-tuned multilingual transformer model trained using Masked Language Modeling on a dataset that includes original, fully transliterated, and partially transliterated text. The fine-tuned embeddings are integrated into a custom transformer classifier designed to capture complex dependencies in multilingual sequences. The system achieves state-of-the-art performance, demonstrating the effectiveness of combining transliteration-aware fine-tuning with robust transformer architectures to handle code-mixed and resource-scarce text, providing a scalable solution for multilingual natural language processing tasks.

This research introduces an innovative Attention BiLSTM-XLM-RoBERTa model for tackling the challenge of fake news detection in Malayalam datasets. By fine-tuning XLM-RoBERTa with Masked Language Modeling (MLM) on transliteration-aware data, the model effectively bridges linguistic and script diversity, seamlessly integrating native, Romanized, and mixed-script text. Although most of the training data is monolingual, the proposed approach demonstrates robust performance in handling diverse script variations. Achieving a macro F1-score of 0.5775 and securing top rankings in the shared task, this work highlights the potential of multilingual models in addressing resource-scarce language challenges and sets a foundation for future advancements in fake news detection.

This research presents a robust multimodal framework for hate speech detection in Malayalam, combining fine-tuned Wav2Vec 2.0, Malayalam-Doc-Topic-BERT, and an Attention-Driven BiLSTM architecture. The proposed approach effectively integrates acoustic and textual features, achieving a macro F1-score of 0.84 on the Malayalam test set. Fine-tuning Wav2Vec 2.0 on Malayalam speech data and leveraging Malayalam-Doc-Topic-BERT significantly improved performance over prior methods using openly available models. The results highlight the potential of language-specific models and advanced multimodal fusion techniques for addressing nuanced hate speech categories, setting the stage for future work on Dravidian languages like Tamil and Telugu.

This study presents a hybrid model integrating TamilXLM-RoBERTa and MalayalamXLM-RoBERTa with BiLSTM and attention mechanisms to classify AI-generated and human-written product reviews in Tamil and Malayalam. The model employs a transliteration-based fine-tuning strategy, effectively handling native, Romanized, and mixed-script text. Despite being trained on a relatively small portion of data, our approach demonstrates strong performance in distinguishing AI-generated content, achieving competitive macro F1 scores in the DravidianLangTech 2025 shared task. The proposed method showcases the effectiveness of multilingual transformers and hybrid architectures in tackling low-resource language challenges.

This research investigates abusive comment detection in Tamil and Malayalam, focusing on code-mixed, multilingual social media text. A hybrid Attention BiLSTM-XLM-RoBERTa model was utilized, combining fine-tuned embeddings, sequential dependencies, and attention mechanisms. Despite computational constraints limiting fine-tuning to a subset of the AI4Bharath dataset, the model achieved competitive macro F1-scores, ranking 6th for both Tamil and Malayalam datasets with minor performance differences. The results emphasize the potential of multilingual transformers and the need for further advancements, particularly in addressing linguistic diversity, transliteration complexity, and computational limitations.

Detecting misogyny in memes is challenging due to their multimodal nature, especially in low-resource languages like Tamil and Malayalam. This paper presents our work in the Misogyny Meme Detection task, utilizing both textual and visual features. We propose an Attention-Driven BiLSTM-XLM-RoBERTa-ResNet model, combining a transliteration-aware fine-tuned XLM-RoBERTa for text analysis and ResNet-50 for image feature extraction. Our model achieved Macro-F1 scores of 0.8805 for Malayalam and 0.8081 for Tamil, demonstrating competitive performance. However, challenges such as class imbalance and domain-specific image representation persist. Our findings highlight the need for better dataset curation, task-specific fine-tuning, and advanced fusion techniques to enhance multimodal hate speech detection in Dravidian languages.

This study investigates sentiment analysis in code-mixed Tamil-English text using an Attention BiLSTM-XLM-RoBERTa model, combining multilingual embeddings with sequential context modeling to enhance classification performance. The model was fine-tuned using masked language modeling and trained with an attention-based BiLSTM classifier to capture sentiment patterns in transliterated and informal text. Despite computational constraints limiting pretraining, the approach achieved a Macro f1 of 0.5036 and ranked first in the competition. The model performed best on the Positive class, while Mixed Feelings and Unknown State showed lower recall due to class imbalance and ambiguity. Error analysis reveals challenges in handling non-standard transliterations, sentiment shifts, and informal language variations in social media text. These findings demonstrate the effectiveness of transformer-based multilingual embeddings and sequential modeling for sentiment classification in code-mixed text.

This paper focuses on detecting fake news in resource-constrained languages, particularly Malayalam. We present a novel framework combining subword tokenization, Sanskrit-transliterated Subword2vec embeddings, and a powerful Bidirectional Long Short-Term Memory (BiLSTM) architecture. Despite using only monolingual Malayalam data, our model excelled in the FakeDetect-Malayalam challenge, ranking 4th. The innovative subword tokenizer achieves a remarkable 200x compression ratio, highlighting its efficiency in minimizing model size without compromising accuracy. Our work facilitates resource-efficient deployment in diverse linguistic landscapes and sparks discussion on the potential of multilingual data augmentation. This research provides a promising avenue for mitigating linguistic challenges in the NLP-driven battle against deceptive content.

This research focuses on Homophobia and Transphobia Detection in Dravidian languages, specifically Telugu, Kannada, Tamil, and Malayalam. Leveraging the Homophobia/ Transphobia Detection dataset, we propose an innovative approach employing a custom-designed tokenizer with a Bidirectional Long Short-Term Memory (BiLSTM) architecture. Our distinctive contribution lies in a tokenizer that reduces model sizes to below 7MB, improving efficiency and addressing real-time deployment challenges. The BiLSTM implementation demonstrates significant enhancements in hate speech detection accuracy, effectively capturing linguistic nuances. Low-size models efficiently alleviate inference challenges, ensuring swift real-time detection and practical deployment. This work pioneers a framework for hate speech detection, providing insights into model size, inference speed, and real-time deployment challenges in combatting online hate speech within Dravidian languages.