Login New user?  
01-Applied Mathematics & Information Sciences
An International Journal


Volumes > Volume 17 > No. 5


Context-Aware Deep Learning Model For Roman Urdu Hate Speech On Social Media Platform

PP: 970-992
M Dhamodaran, R Ranjithkumar, G Abishek, AS Aabidkaleem2, S Dhinakaran,
Through social media applications like Twitter and Facebook, users can communicate and share their thoughts, status updates, opinions, photographs, and videos with the rest of the world. Unfortunately, some people use these platforms to spread hate speech and abusive language. The improvement of scorn talk could achieve scorn infringement, advanced brutality, and huge wickedness to the web, real security, and social prosperity. An expert team used the annotation rules to create a brand-new Roman Urdu Hate Speech Dataset (RU-HSD-30K) from the existing data. The Bi-LSTM model with an attention layer for Roman-Urdu Hate Speech Detection has not been investigated to our knowledge. Hence, existing fostered a setting mindful Roman Urdu Disdain Discourse identification model in view of Bi-LSTM with a consideration layer and utilized custom word2vec for word embeddings. It inspected the impact of lexical standardization of Roman Urdu words on the exhibition of the proposed model. Different conventional as well as profound learning models, including LSTM and CNN models, were utilized as standard models. The exhibition of the models was surveyed as far as assessment networks like exactness, accuracy, review, and F1-score. A cross-domain dataset is used to evaluate each model's generalizability. Trial results uncovered that Bi-LSTM with consideration beat the conventional AI models and other profound learning models with a precision score of 0.875 and a F-Score of 0.885. In addition, the findings demonstrated that when applied to unseen data, our suggested model—Bi-LSTM with Attention Layer—is more comprehensive than previous models. The findings demonstrated that the proposed model performed better when Roman Urdu words were lexically normalized. Due to its ability to capture the context of the text, the proposed system is used as a transformer-based model for grouping Roman Urdu disdain discourse. The presentation of each model was assessed with regards to exactness, accuracy, review, and Fmeasure. A cross-domain dataset served as the basis for evaluating each model's generalizability. When directly applied to the classification task of Roman Urdu hate speech, the experimental results showed that the transformer-based model performed better than traditional machine learning, deep learning models using DNN, and pre-trained transformer- based models in terms of accuracy, precision, recall, and F-measure, with scores of 96.74 percent and 97.89 percent, respectively. On a cross-domain dataset, the transformer-based model also demonstrated superior generalization.

  Home   About us   News   Journals   Conferences Contact us