Scholarlink

While the rapid expansion of social networks has revolutionised communication, it has also given rise to the pervasive problem of toxic comments, which harm people and upend online groups. Due to linguistic complexity, a lack of annotated datasets, and the lack of pre-trained models, it can be particularly difficult to detect such comments in low-resource languages like Assamese. To overcome these obstacles, this study presents hybrid deep learning models, namely LSTM + BiLSTM and BiLSTM + LSTM. Preprocessing methods, such as tokenisation, lemmatisation, and data augmentation with Assamese WordNet, were used to guarantee data quality and balance using a curated dataset of 100,000 Assamese comments. Linguistic aspects were represented via word embeddings, which allowed models to identify contextual and semantic patterns. With 90% accuracy and balanced F1-scores of 0.90 and 0.91%, the LSTM + BiLSTM model outperformed the BiLSTM + LSTM model, which had 88% accuracy. In addressing linguistic idiosyncrasies, enhancing toxic comment identification for low-resource languages, and laying out the basis for research to follow, the results show how effective hybrid techniques are.

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

Combining Bidirectional and Sequential Neural Networks for Assamese Toxic Comment Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mandira Neog,
  • Nomi Baruah

摘要

While the rapid expansion of social networks has revolutionised communication, it has also given rise to the pervasive problem of toxic comments, which harm people and upend online groups. Due to linguistic complexity, a lack of annotated datasets, and the lack of pre-trained models, it can be particularly difficult to detect such comments in low-resource languages like Assamese. To overcome these obstacles, this study presents hybrid deep learning models, namely LSTM + BiLSTM and BiLSTM + LSTM. Preprocessing methods, such as tokenisation, lemmatisation, and data augmentation with Assamese WordNet, were used to guarantee data quality and balance using a curated dataset of 100,000 Assamese comments. Linguistic aspects were represented via word embeddings, which allowed models to identify contextual and semantic patterns. With 90% accuracy and balanced F1-scores of 0.90 and 0.91%, the LSTM + BiLSTM model outperformed the BiLSTM + LSTM model, which had 88% accuracy. In addressing linguistic idiosyncrasies, enhancing toxic comment identification for low-resource languages, and laying out the basis for research to follow, the results show how effective hybrid techniques are.