TL;DR
This paper tackles the critical issue of gender-based online abuse targeting women in Tamil and Malayalam. Researchers developed and compared machine learning models (SVM, Logistic Regression, Naive Bayes) to classify abusive comments, achieving a best performance of 89.89% accuracy and a 90% F1 score with SVM. This work contributes to building more robust content moderation systems for low-resource languages.
What Problem Does This Paper Address?
Social media has become a double-edged sword: a powerful tool for connection and information sharing, yet also a breeding ground for cyberbullying and harassment, disproportionately affecting women. This abuse often manifests as degrading comments rooted in deep-seated cultural prejudices and gender inequality. While detecting such abuse in high-resource languages is challenging, it's even more so in low-resource languages like Tamil and Malayalam. These languages often feature implicit bias, stereotypes, and coded language, making them difficult for automated systems to detect. This paper aims to bridge this gap by developing effective content moderation tools for these specific languages.
Methods
The researchers employed a multi-stage approach:
- Preprocessing: This involved standard text cleaning techniques such as converting text to lowercase, stripping punctuation, removing numbers, and lemmatization to reduce words to their root forms while preserving meaning.
- Feature Engineering: They utilized Term Frequency-Inverse Document Frequency (TF-IDF) to convert text into numerical features. TF-IDF is a common technique that weighs words based on their frequency in a document and their rarity across a corpus, helping to identify important terms. They specifically used unigrams (single words) and bigrams (two-word sequences) to capture contextual relationships.
- Machine Learning Models: Three classical machine learning models were implemented and compared:
- Logistic Regression: A probabilistic model that predicts the likelihood of a comment being abusive.
- Support Vector Machine (SVM): A model that finds an optimal hyperplane to separate abusive from non-abusive comments in the TF-IDF feature space.
- Naive Bayes: A probabilistic classifier that uses word frequencies to determine the likelihood of a comment belonging to a particular class.
- Hyperparameter Tuning: GridSearchCV was used to tune the hyperparameters of these models to optimize their performance.
Datasets
The study utilized YouTube comments in Tamil and Malayalam. The datasets were adapted and modified from publicly accessible datasets originally published as part of the DravidianLangTech@NAACL2025 program. Each language's dataset was divided into training, validation, and test sets, with comments labeled as either "Abusive" or "Non-Abusive."
Key Insights and Ideas
- SVM's Superiority in Low-Resource Languages: The study highlights that even with classical machine learning models and TF-IDF features, SVM can achieve robust performance in detecting abusive content in low-resource languages like Tamil and Malayalam.
- Importance of TF-IDF for Context: The use of TF-IDF, especially with bigrams, proved effective in capturing contextual relationships within comments, which is crucial for identifying nuanced forms of abuse.
- WordCloud for Interpretability: The use of WordCloud for exploratory data analysis provided valuable insights into common abusive and non-abusive terms, aiding in the interpretation of language patterns specific to these languages.
- Challenges of Implicit Abuse: The paper acknowledges that while their models perform well, they struggle with deeply implicit abuse, sarcasm, and code-mixed language, indicating areas for future improvement.
Key Results
The models were evaluated using accuracy, F1-score, precision, and recall.
- SVM achieved the highest performance with an accuracy of 89.89% and a balanced F1 score of 90%.
- Logistic Regression achieved an accuracy of 87.48% and an F1 score of 87%.
- Naive Bayes achieved an accuracy of 87.51% and an F1 score of 87%.
The ROC curve analysis also indicated SVM as the top performer, followed by Logistic Regression, with Naive Bayes showing relatively poorer performance.
Limitations and Future Work
The authors acknowledge several limitations:
- Lack of Deep Contextual Understanding: The current models, based on TF-IDF, do not possess deep contextual understanding and may fail to incorporate implicit abuse, sarcasm, or nuanced conversational elements.
- Multimodal Abuse: The system currently only processes text and cannot detect abuse embedded in images, memes, or other multimedia formats.
- Limited Linguistic Breadth: The training data's linguistic and cultural breadth might limit the model's generalization to dialectic and other variant forms of spoken Tamil and Malayalam.
- Bias in Data: Biased training data can lead to unfair classification, negatively impacting specific user groups.
- Computational Demands: Real-time deployment of deep learning models poses computational challenges.
Future work proposed includes:
- Upgraded Transformer Models: Utilizing advanced transformer-based models like BERT, mBERT, and IndicBERT for improved contextual understanding and classification.
- Multimodal Abuse Detection: Integrating modules to analyze textual and visual aspects of content, especially in memes and multimedia.
- Data Augmentation: Expanding datasets with variations in linguistic and cultural differences.
- Model Optimization: Employing techniques like quantization, pruning, and knowledge distillation to reduce computational overhead.
- Explainable AI (XAI): Applying XAI strategies for bias mitigation and to make the system fairer and more interpretable.
- Cross-Lingual Transfer Learning: Extending the system to support multiple Dravidian languages.
Summary updated