Scholarlink

<p>The demand for explainable AI in misinformation detection is crucial for building user trust and understanding model behavior. Many recent methods try to explain how they spot fake news using text, images, or both (multimodal). However, these methods often rely on fixed-size explanations (text and images) generated through ranking-based systems, which fail to effectively differentiate between explainable and non-explainable components. This shortcoming results in vague explanations and limited model performance. To overcome these aforesaid issues, we come up with a <Emphasis Type="Underline">m</Emphasis>ultimodal <Emphasis Type="Underline">EX</Emphasis>plainable misinformation detection method based on <Emphasis Type="Underline">AC</Emphasis>ute <Emphasis Type="Underline">T</Emphasis>hresholding mechanism (<i>mEXACT</i>) that identifies a variable-size bucket of check-worthy information, when removed, can flip the model’s prediction from fake to real. Identifying minimal set of significant information enables our model to distinguish between contributing and non-contributing misinformation components, thereby enhancing interpretability while improving classification performance. Extensive experiments on two real-world multimodal COVID-19 misinformation datasets, ReCOVery and MMCoVaR, demonstrate that mEXACT significantly outperforms state-of-the-art techniques, achieving <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10489_2025_6656_Article_IEq1.gif" Format="GIF" Height="19" Rendition="HTML" Resolution="72" Type="Linedraw" Width="89" /> </InlineMediaObject> <EquationSource Format="TEX">\((6.8-8.7)\%\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10489_2025_6656_Article_IEq2.gif" Format="GIF" Height="19" Rendition="HTML" Resolution="72" Type="Linedraw" Width="89" /> </InlineMediaObject> <EquationSource Format="TEX">\((4.9-5.4)\%\)</EquationSource> </InlineEquation> higher Accuracy-F1 scores, respectively.</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

Beyond just saying it’s false: explainable AI for multimodal misinformation detection

  • 期刊论文

  • 发表日期：

    2025年6月19日

查看全文
  • Saswata Roy,
  • Manish Bhanu,
  • Shalini Priya,
  • Joydeep Chandra,
  • Sourav Kumar Dandapat

摘要

The demand for explainable AI in misinformation detection is crucial for building user trust and understanding model behavior. Many recent methods try to explain how they spot fake news using text, images, or both (multimodal). However, these methods often rely on fixed-size explanations (text and images) generated through ranking-based systems, which fail to effectively differentiate between explainable and non-explainable components. This shortcoming results in vague explanations and limited model performance. To overcome these aforesaid issues, we come up with a multimodal EXplainable misinformation detection method based on ACute Thresholding mechanism (mEXACT) that identifies a variable-size bucket of check-worthy information, when removed, can flip the model’s prediction from fake to real. Identifying minimal set of significant information enables our model to distinguish between contributing and non-contributing misinformation components, thereby enhancing interpretability while improving classification performance. Extensive experiments on two real-world multimodal COVID-19 misinformation datasets, ReCOVery and MMCoVaR, demonstrate that mEXACT significantly outperforms state-of-the-art techniques, achieving \((6.8-8.7)\%\) and \((4.9-5.4)\%\) higher Accuracy-F1 scores, respectively.