United States

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.

NAACL 2025

byteSizedLLM@DravidianLangTech 2025: Abusive Tamil and Malayalam Text targeting Women on Social Media Using XLM-RoBERTa and Attention-BiLSTM

attention bilstm-xlm-roberta

multilingual social media

abusive comment detection

code-mixed text

### Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

Welcome to the 2025 meeting of the Nations of the Americas Chapter of the Association for Computational Linguistics! I am proud to help organize the first NAACL conference to carry the new name of our organization, one that emphasizes inclusion for all of the Americas. I am also pleased to welcome you to Albuquerque, New Mexico, a state whose unique blend of cultural influences will make for an excellent backdrop for NAACL 2025, especially with this year’s special theme on NLP in a Multicultural World. 
**[Continue reading...](https://drive.google.com/file/d/1jX-qGhqVSZZCIrAnJaz798pflu5Irrdn/view?usp=sharing)**

*- Colin Cherry, Google, NAACL 2025 General Chair* 

[![](https://assets.underline.io/markdown_image/1/image/b087f8a4dc5816d6e1a6514e59c59ac3.png)](https://drive.google.com/file/d/1T96GzPqObXrMln2BMByCSXTSizjTg69P/view?usp=sharing)

You need to log in with the email address you registered with. Access credentials have been sent to your email. 

Please be sure to check your spam and other email folders if you do not see an email confirmation right away.

Please log in to explore this event.

To access NAACL 2025 event page you are required to register. Please follow [**this link**](https://2025.naacl.org/registration/registration/) to register. Access will depend on your registration type.

Please register!

Welcome to 2025 Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics

workshop paper

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.

byteSizedLLM@DravidianLangTech 2025: Multimodal Misogyny Meme Detection in Low-Resource Dravidian Languages Using Transliteration-Aware XLM-RoBERTa, ResNet-50, and Attention-BiLSTM

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.

byteSizedLLM@DravidianLangTech 2025: Sentiment Analysis in Tamil Using Transliteration-Aware XLM-RoBERTa and Attention-BiLSTM

Hate speech detection is a serious challenge due to the different digital media communication, particularly in low-resource languages. This research focuses on the problem of multimodal hate speech detection by incorporating both textual and audio modalities. In the context of social media platforms, hate speech is conveyed not only through text but also through audios, which may further amplify harmful content. In order to manage the issue, we provide a multiclass classification model that influences both text and audio features to detect and categorize hate speech in low-resource languages. The model uses machine learning models for text analysis and audio processing, allowing it to efficiently capture the complex relationships between the two modalities. Class weight mechanism involves avoiding overfitting. The prediction has been finalized using the majority fusion technique. Performance is measured using a macro average F1 score metric. Three languages—Tamil, Malayalam, and Telugu—have the optimal F1-scores, which are 0.59, 0.52, and 0.33.

SSNCSE@DravidianLangTech 2025: Multimodal Hate Speech Detection in Dravidian Languages

Detecting fake news in under-resourced Dravidian languages is a rigorous task due to the scarcity of annotated datasets and the intricate nature of code-mixed text. This study tackles these issues by employing advanced machine learning techniques for two key classification tasks, the first task involves binary classification achieving a macro-average F1-score of 0.792 using a hybrid fusion model that integrates Bidirectional Recurrent Neural Network (Bi-RNN) and Long Short-Term Memory (LSTM)-Recurrent Neural Network (RNN) with weighted averaging. The second task focuses on fine-grained classification, categorizing news where an LSTM-GRU hybrid model attained a macro-average F1-score of 0.26. These findings highlight the effectiveness of hybrid models in improving fake news detection for under-resourced languages. Additionally, this study provides a foundational framework that can be adapted to address similar challenges in other under-resourced languages, emphasizing the need for further research in this area.

YenCS@DravidianLangTech 2025: Integrating Hybrid Architectures for Fake News Detection in Low-Resource Dravidian Languages

The increasing prevalence of AI-generated content, including fake product reviews, poses significant challenges in maintaining authenticity and trust in e-commerce systems. While much work has focused on detecting such reviews in high-resource languages, limited attention has been given to low-resource languages like Malayalam and Tamil. This study aims to address this gap by developing a robust framework to identify AI-generated product reviews in these languages. We explore a BERT-based approach for this task. Our methodology involves fine-tuning a BERT-based model specifically on Malayalam and Tamil datasets. The experiments are conducted using labeled datasets that contain a mix of human-written and AI-generated reviews. Performance is evaluated using the macro F1 score. The results show that the BERT-based model achieved a macro F1 score of 0.6394 for Tamil and 0.8849 for Malayalam. Preliminary results indicate that the BERT-based model performs significantly better for Malayalam than for Tamil in terms of the average Macro F1 score, leveraging its ability to capture the complex linguistic features of these languages. Finally, we open the source code of the implementation in the GitHub repository: AI-Generated-Product-Review-Code

Girma@DravidianLangTech 2025: Detecting AI Generated Product Reviews

Research on political feeling is essential for comprehending public opinion in the digital age, as social media and news platforms are often the sites of discussions. To categorize political remarks into sentiments like positive, negative, neutral, opinionated, substantiated, and sarcastic, this study offers a multiclass sentiment analysis approach. We trained models, such as Random Forest and a Feedforward Neural Network, after preprocessing and feature extraction from a large dataset of political texts using Natural Language Processing approaches. The Random Forest model, which was great at identifying more complex attitudes like sar casm and opinionated utterances, had the great est accuracy of 84%, followed closely by the Feedforward Neural Network model, which had 83%. These results highlight how well political discourse can be analyzed by combining deep learning and traditional machine learning techniques. There is also room for improvement by adding external metadata and using sophisticated models like BERT for better sentiment classification.

Beyond_Tech@DravidianLangTech 2025: Political Multiclass Sentiment Analysis using Machine Learning and Neural Network

Detecting abusive and similarly toxic content posted on a social media platform is challenging due to the complexities of the language, data imbalance, and the code-mixed nature of the text. In this paper, we present our submissions for the shared task on abusive Tamil and Malayalam texts targeting women on social media—DravidianLangTech@NAACL 2025. We propose a hybrid embedding model that integrates embeddings generated using term frequency-inverse document frequency (TF-IDF) and BERT. To get rid of the differences in the embedding dimensions, we used a dimensionality reduction method with TF-IDF embedding. We submitted two more runs to the shared task, which involve a model based on TF-IDF embedding and another based on BERT-based embedding. The code for the submissions is available at https://github.com/Tarrruh/NLP_HTMS.

HTMS@DravidianLangTech 2025: Fusing TF-IDF and BERT with Dimensionality Reduction for Abusive Language Detection in Tamil and Malayalam

This work proposed a methodology for assessing political sentiments in Tamil tweets using machine learning models. The approach addressed linguistic challenges in Tamil text, including cleaning, normalization, tokenization, and class imbalance, through a robust preprocessing pipeline. Various models, including Random Forest, Logistic Regression, and CatBoost, were applied, with Random Forest achieving a macro F1-score of 0.2933 and securing 8th rank among 153 participants in the Codalab competition. This accomplishment highlights the effectiveness of machine learning models in handling the complexities of multilingual, code-mixed, and unstructured data in Tamil political discourse. The study also emphasized the importance of tailored preprocessing techniques to improve model accuracy and performance. It demonstrated the potential of computational linguistics and machine learning in understanding political discourse in low-resource languages like Tamil, contributing to advancements in regional sentiment analysis.

Team_Catalysts@DravidianLangTech 2025: Leveraging Political Sentiment Analysis using Machine Learning Techniques for Classifying Tamil Tweets

The rise of misogynistic memes on social media posed challenges to civil discourse. This paper aimed to detect misogyny in Dravidian language memes using a multimodal deep learning approach. We integrated Bidirectional Encoder Representations from Transformers (BERT), Long Short-Term Memory (LSTM), EfficientNet, and a Vision Language Model (VLM) to analyze textual and visual informa tion. EfficientNet extracted image features, LSTM captured sequential text patterns, and BERT learned language-specific embeddings. Among these, VLM achieved the highest accuracy of 85.0% and an F1-score of 70.8, effectively capturing visual-textual relationships. Validated on a curated dataset, our method outperformed baselines in precision, recall, and F1-score. Our approach ranked 12th out of 118 participants for the Tamil language, highlighting its competitive performance. This research emphasizes the importance of multimodal models in detecting harmful content. Future work can explore improved feature fusion techniques to enhance classification accuracy.

InnovationEngineers@DravidianLangTech 2025: Enhanced CNN Models for Detecting Misogyny in Tamil Memes Using Image and Text Classification

The detection of abusive text targeting women, especially in Dravidian languages like Tamil and Malayalam, presents a unique challenge due to linguistic complexities and code-mixing on social media. This paper evaluates machine learning models such as Support Vector Machines (SVM), Logistic Regression (LR), and Random Forest Classifiers (RFC) for identifying abusive content. Code-mixed datasets sourced from platforms like YouTube are used to train and test the models. Performance is evaluated using accuracy, precision, recall, and F1-score metrics. Our findings show that SVM outperforms the other classifiers in accuracy and recall. However, challenges persist in detecting implicit abuse and addressing informal, culturally nuanced language. Future work will explore transformer-based models like BERT for better context understanding, along with data augmentation techniques to enhance model performance. Additionally, efforts will focus on expanding labeled datasets to improve abuse detection in these low-resource languages.

Downloads

Next from NAACL 2025

byteSizedLLM@DravidianLangTech 2025: Multimodal Misogyny Meme Detection in Low-Resource Dravidian Languages Using Transliteration-Aware XLM-RoBERTa, ResNet-50, and Attention-BiLSTM

Similar lecture

MSM_CUET@DravidianLangTech 2025: XLM-BERT and MuRIL Based Transformer Models for Detection of Abusive Tamil and Malayalam Text Targeting Women on Social Media