Scholarlink

There is a preponderance of AI-generated text everywhere today. Literature shows that there has been considerable success in detecting such text. This paper uses explainable AI (XAI) techniques to get insights into the workings of the machine learning models used to classify synthetic text. In detecting such text, this work analyzes synthetic and genuine information from visualization and explainability perspectives. The text is converted into vector embeddings using Robustly Optimized BERT Pretraining Approach (RoBERTa). Variational Autoencoders (VAEs) are used for visualization and Support Vector Machine (SVM) is used for classification. Local Interpretable Model-Agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), and Integrated Gradients are used to explain the classification. The experiments are done on two different types of datasets. Despite the machine learning model achieving outstanding accuracy similar to the previous work in the literature, it was determined that there is no clear explanation of why the representation learning or the classification works so outstandingly well. The explainability techniques used show that the model focuses on words that do not clearly indicate that the text is synthetically generated. Visualization in two dimensions shows that the vector embeddings of both classes of text overlap significantly and that there is no clear separation of the representations learned.

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

Unraveling the Enigma of Classification of Synthetic and Genuine Information Using Machine Learning and Explainable AI

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Vishnu S. Pendyala

摘要

There is a preponderance of AI-generated text everywhere today. Literature shows that there has been considerable success in detecting such text. This paper uses explainable AI (XAI) techniques to get insights into the workings of the machine learning models used to classify synthetic text. In detecting such text, this work analyzes synthetic and genuine information from visualization and explainability perspectives. The text is converted into vector embeddings using Robustly Optimized BERT Pretraining Approach (RoBERTa). Variational Autoencoders (VAEs) are used for visualization and Support Vector Machine (SVM) is used for classification. Local Interpretable Model-Agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), and Integrated Gradients are used to explain the classification. The experiments are done on two different types of datasets. Despite the machine learning model achieving outstanding accuracy similar to the previous work in the literature, it was determined that there is no clear explanation of why the representation learning or the classification works so outstandingly well. The explainability techniques used show that the model focuses on words that do not clearly indicate that the text is synthetically generated. Visualization in two dimensions shows that the vector embeddings of both classes of text overlap significantly and that there is no clear separation of the representations learned.