Tools
Tools
Abdul Kareem, Syed Ebrahim (2023) Leveraging Transfer Learning Techniques for Homophobia and Transphobia Detection. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (881kB) | Preview |
Preview |
PDF (Configuration manual)
Download (2MB) | Preview |
Abstract
The proliferation of hate speech and abusive content on social media makes it unsafe place for networking. Unlike past, discussion on LGBTQ+ topic is openly happening on public forums at present. People have started to openly declare their sexual orientation and gender identification. Many nations have legalized such acts which were once considered a crime or sin. Due to anonymity and aversion on LGBTQ+ community, some users spread hatred towards them on social media. To ensure equality, inclusiveness and diversity, social media companies must detect and moderate such posts. This will ensure harmony and maintain decorum on their platforms. In Natural Language Processing domain, one of the most active areas of research is text classification. Over the last few years, Transformer based models are widely used for this task as they provide better results than conventional Machine Learning models. In this research, RoBERTa, DistilBERT and mBERT were implemented and their performance is compared with past works. The results shows that DistilBERT outperformed RoBERTa with a macro avg F1 score of 0.47 on English dataset. This is slightly better than conventional ML models like Logistic Regression and SVM. On the Tamil dataset, mBERT model achieved a macro avg F1 score of 0.83 which is 3.75% improvement over previous studies.
Actions (login required)
 |
View Item |