The increase in abusive content on online social media platforms is impacting the social life of online users. Use of offensive and hate speech has been making so-cial media toxic. Homophobia and transphobia constitute offensive comments against LGBT+ community. It becomes imperative to detect and handle these comments, to timely flag or issue a warning to users indulging in such behaviour. However, automated detection of such content is a challenging task, more so in Dravidian languages which are identified as low resource languages. Motivated by this, the paper attempts to explore applicability of different deep learning mod-els for classification of the social media comments in Malayalam and Tamil lan-guages as homophobic, transphobic and non-anti-LGBT+content. The popularly used deep learning models- Convolutional Neural Network (CNN), Long Short Term Memory (LSTM) using GloVe embedding and transformer-based learning models (Multilingual BERT and IndicBERT) are applied to the classification problem. Results obtained show that IndicBERT outperforms the other imple-mented models, with obtained weighted average F1-score of 0.86 and 0.77 for Malayalam and Tamil, respectively. Therefore, the present work confirms higher performance of IndicBERT on the given task in selected Dravidian languages.

翻译：在线社交媒体平台中滥用内容日益增多，正影响着网络用户的社会生活。攻击性言论和仇恨言论的使用已使社交媒体环境趋于恶化。恐同和恐跨性别言论是针对LGBT+群体的攻击性评论。及时检测并处理这些评论，对涉及此类行为的用户进行标记或发出警告，已成为当务之急。然而，自动化检测此类内容是一项具有挑战性的任务，尤其是在被认定为低资源语言的达罗毗荼语系中。受此启发，本文尝试探索不同深度学习模型在马拉雅拉姆语和泰米尔语社交媒体评论分类中的应用，将其分为恐同、恐跨性别及非反LGBT+内容三类。研究采用了主流的深度学习模型——使用GloVe词嵌入的卷积神经网络（CNN）、长短期记忆网络（LSTM），以及基于Transformer的学习模型（多语言BERT和IndicBERT）来解决分类问题。结果表明，IndicBERT的性能优于其他已实现模型，在马拉雅拉姆语和泰米尔语上分别取得了0.86和0.77的加权平均F1分数。因此，本研究证实了IndicBERT在所选达罗毗荼语系特定任务中的更高性能。