Abstract:
There is an immediate demand for automated systems that can identify and handle hostile
content due to the quick spread of online communication platforms. This study tackles the
problem of multi-label hostility categorization in the context of the Bangla language, where
text texts may concurrently include several forms of hostile content. This study offers a
thorough method for dealing with the classification of multi-label antagonism in Bangla
literature. A deep learning model with various types of hostility is constructed and trained
on a large-scale annotated dataset of Bangla text using the latest advances in natural
language processing techniques. To capture complex linguistic patterns and contextual
connections in Bangla text, the model design combines attention processes with
sophisticated neural network components including recurrent and convolutional layers.
Additionally, this study looks into a number of pre-processing methods designed especially
for the Bangla language, such as addressing typographical variances, word segmentation,
and stemming. Furthermore, the study investigates feature engineering techniques to
improve the model's capacity to identify subtle language cues that convey animosity in
various settings. Experiments carried out on benchmark datasets show how reliable and
efficient the suggested method is in correctly identifying various forms of offensive content
in Bangla text. We achieved the most optimal result at a training accuracy of 87.67%, and
a training loss of 36.89% by using Bi-LSTM model. The model demonstrates its potential
utility in practical applications such as sentiment analysis in Bangla-speaking
communities, online safety, and content moderation by achieving competitive performance
metrics in precision, recall, and F1-score. Ultimately, by providing insights and approaches
that may be modified and expanded to handle comparable problems in other languages and
cultural situations, this research advances automated hostility detection systems for the
Bangla language.
