Scholarlink

<p>In recent years, the exchange of offensive content through social media has increased. Social media users represent their thoughts in regional language. Such comments contain mixing of more than one language. There is no standard tool or script for effectively identifying and detecting these code-mixed comments. In general social media comments may be lengthy and the long term dependency need to be captured. The proportion of offensive comments posted in social media is very less, there by creating an imbalanced data distribution. To address these limitations, Multi Channel CNN is proposed that extracts n-gram features, which are later concatenated by early fusion method. The obtained semantically rich features are balanced in the feature space by applying SOUP based sampling technique. These features are fed to XGB classifier for classifying the YouTube comments as Offensive or not. To prove the effectiveness of the proposed model, we have conducted extensive experiments namely Unidirectional, Bidirectional, Hybrid, Multi-Channel Models and Transfer learning based Multi-Channel models on YouTube Offensive Comments in two different languages, Malayalam and Kannada. The proposed Transfer learning based Multichannel CNN model outperformed other methods by achieving an F1 score of 97.7%, 72.24% on Malayalam and Kannada languages. The proposed approach demonstrates notable performance gain by 3% in Malayalam and 11% in Kannada over existing state-of-the-art methods.</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

Code-mixed offensive content categorization using multi-channel CNN with N-gram fusion and XGB

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • R. Ramesh Kannan,
  • Ratnavel Rajalakshmi

摘要

In recent years, the exchange of offensive content through social media has increased. Social media users represent their thoughts in regional language. Such comments contain mixing of more than one language. There is no standard tool or script for effectively identifying and detecting these code-mixed comments. In general social media comments may be lengthy and the long term dependency need to be captured. The proportion of offensive comments posted in social media is very less, there by creating an imbalanced data distribution. To address these limitations, Multi Channel CNN is proposed that extracts n-gram features, which are later concatenated by early fusion method. The obtained semantically rich features are balanced in the feature space by applying SOUP based sampling technique. These features are fed to XGB classifier for classifying the YouTube comments as Offensive or not. To prove the effectiveness of the proposed model, we have conducted extensive experiments namely Unidirectional, Bidirectional, Hybrid, Multi-Channel Models and Transfer learning based Multi-Channel models on YouTube Offensive Comments in two different languages, Malayalam and Kannada. The proposed Transfer learning based Multichannel CNN model outperformed other methods by achieving an F1 score of 97.7%, 72.24% on Malayalam and Kannada languages. The proposed approach demonstrates notable performance gain by 3% in Malayalam and 11% in Kannada over existing state-of-the-art methods.