Scholarlink

<p>The widespread use of social media has brought people from all over the world closer, but it has also become a platform for spreading hate, offensive content, and misinformation. Various forms of media, such as text, images, audio, and video, are used to disseminate information. In this paper, we tackle the task of identifying offensive and non-offensive text in the context of women’s harassment, considering the intensity of the offense. We propose a novel approach that leverages sentiment and emotion, including their intensity, to enhance the performance of this task. To facilitate our research, we constructed a large-scale dataset called the Women and Child-related Sentiment and Emotion-Aware offensive dataset (WCSEoff) to aid in identifying offensive content. Furthermore, we develop a multi-task model that not only detects whether images contain offensive or non-offensive content and quantifies the intensity of offensiveness as the primary task but also classifies the sentiment and emotions present, along with their respective intensities, as the auxiliary task. By adopting this multitasking framework, we achieve a performance improvement of over 7% compared to approaches that do not incorporate sentiment and emotion analysis. Code and data will be available at <a href="https://www.iitp.ac.in/%7eai-nlp-ml/resources.html#MOSAIC">https://www.iitp.ac.in/~ai-nlp-ml/resources.html#MOSAIC</a>.</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

Unmasking offensive content: a multimodal approach with emotional understanding

  • 期刊论文

  • 发表日期：

    2025年1月7日

查看全文
  • Gopendra Vikram Singh,
  • Soumitra Ghosh,
  • Mauajama Firdaus,
  • Asif Ekbal,
  • Pushpak Bhattacharyya

摘要

The widespread use of social media has brought people from all over the world closer, but it has also become a platform for spreading hate, offensive content, and misinformation. Various forms of media, such as text, images, audio, and video, are used to disseminate information. In this paper, we tackle the task of identifying offensive and non-offensive text in the context of women’s harassment, considering the intensity of the offense. We propose a novel approach that leverages sentiment and emotion, including their intensity, to enhance the performance of this task. To facilitate our research, we constructed a large-scale dataset called the Women and Child-related Sentiment and Emotion-Aware offensive dataset (WCSEoff) to aid in identifying offensive content. Furthermore, we develop a multi-task model that not only detects whether images contain offensive or non-offensive content and quantifies the intensity of offensiveness as the primary task but also classifies the sentiment and emotions present, along with their respective intensities, as the auxiliary task. By adopting this multitasking framework, we achieve a performance improvement of over 7% compared to approaches that do not incorporate sentiment and emotion analysis. Code and data will be available at https://www.iitp.ac.in/~ai-nlp-ml/resources.html#MOSAIC.