Scholarlink

<p>The spread of hate speech at such a fast pace on the social media networks presents serious detriment to the digital security, psychological well being, social cohesion. Moderation of such dangerous contents cannot take place manually because online activity is extremely large and is continuously changing. This study showcases an optimized Bidirectional Long Short Term Memory (BiLSTM) machine with the use of FastText based feature extraction to classify and identify cyberbullying, and hate speech through a multi class. This dataset originates from Twitter is categorized into three classes: ethnicity based hate speech, other forms of cyberbullying, and non cyberbullying content. Our research paper involve 3 steps data preprocessing, feature data preprocessing followed by feature extraction and classification techniques, include tokenization, normalization, and stopword removal, were applied to improve data quality.In feature extraction we use FastText method for embeddings effectively captured subword level information, handling misspellings and slang typical of social media text, in classification proposed optimized BiLSTM model, optimized using the Adam optimizer and fine tuned hyperparameters. The optimized BiLSTM achieved an impressive 94% accuracy in the experiments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimized BiLSTM for Multi-Class Hate Speech Detection with Improved Performance

  • 期刊论文

  • 发表日期：

    2026年4月4日

查看全文
  • Pragya Goswami,
  • A. Daniel

摘要

The spread of hate speech at such a fast pace on the social media networks presents serious detriment to the digital security, psychological well being, social cohesion. Moderation of such dangerous contents cannot take place manually because online activity is extremely large and is continuously changing. This study showcases an optimized Bidirectional Long Short Term Memory (BiLSTM) machine with the use of FastText based feature extraction to classify and identify cyberbullying, and hate speech through a multi class. This dataset originates from Twitter is categorized into three classes: ethnicity based hate speech, other forms of cyberbullying, and non cyberbullying content. Our research paper involve 3 steps data preprocessing, feature data preprocessing followed by feature extraction and classification techniques, include tokenization, normalization, and stopword removal, were applied to improve data quality.In feature extraction we use FastText method for embeddings effectively captured subword level information, handling misspellings and slang typical of social media text, in classification proposed optimized BiLSTM model, optimized using the Adam optimizer and fine tuned hyperparameters. The optimized BiLSTM achieved an impressive 94% accuracy in the experiments.