Scholarlink

In recent years, the rise of AI-driven tools has threatened academic integrity, with students increasingly relying on AI-generated content and compromising their scholarly efforts. To address the challenge of distinguishing machine-generated text from human-written content, this study utilizes the dataset provided by SemEval-2024. We explore various methodologies, comparing traditional machine learning (ML) techniques enhanced with natural language preprocessing (NLP) against the fine-tuning of Large Language Models (LLMs) like XLM-RoBERTa for text classification. Through systematic evaluation across SemEval-2024’s training, development, and testing sets, we uncover key insights. While traditional ML methods with NLP refinements offer a pragmatic approach, transformer models, particularly XLM-RoBERTa, demonstrate superior efficacy. Our findings underscore the transformative potential of transformer models in detecting machine-generated texts, especially in multilingual environments. Additionally, we highlight the effectiveness of majority voting mechanisms in enhancing precision in multilingual contexts.

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

Transfer Learning to Identify Multilingual Machine Genenrated Text

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Samyak Jain,
  • Anshul Bhardwaj,
  • Mohit Somani,
  • Alka Kumari,
  • Disha Goyal,
  • Pritish Kumar Varadwaj

摘要

In recent years, the rise of AI-driven tools has threatened academic integrity, with students increasingly relying on AI-generated content and compromising their scholarly efforts. To address the challenge of distinguishing machine-generated text from human-written content, this study utilizes the dataset provided by SemEval-2024. We explore various methodologies, comparing traditional machine learning (ML) techniques enhanced with natural language preprocessing (NLP) against the fine-tuning of Large Language Models (LLMs) like XLM-RoBERTa for text classification. Through systematic evaluation across SemEval-2024’s training, development, and testing sets, we uncover key insights. While traditional ML methods with NLP refinements offer a pragmatic approach, transformer models, particularly XLM-RoBERTa, demonstrate superior efficacy. Our findings underscore the transformative potential of transformer models in detecting machine-generated texts, especially in multilingual environments. Additionally, we highlight the effectiveness of majority voting mechanisms in enhancing precision in multilingual contexts.