Scholarlink

<p>Approaches constructed using GPT have enormous promise in financial applications in the time of rapidly developing NLP. However, their integration with financial datasets is still at odds, especially when it comes to evaluating the effectiveness and applicability of such models. To handle these problems, this paper proposes a novel bias detection and mitigation framework powered by an Adaptive Deep Crested porcupine Graph neural Network (ADCGN). It leverages graph neural networks’ powerful ability to learn complex relations in multimodal data-textual and numerical inputs-to extensive bias analysis. It incorporates three major stages: pre-processing through BERT-based MobileNetV3 for cleaning the data and extracting relevant features, in-processing through the proposed module for detecting bias, and postprocessing to mitigate the biases in large language models. It was then experimentally verified using FinGPT-a financial chatbot-showing very superior performance across fairness metrics, reducing bias with improvements in accuracy of 94.56%, precision of 92.47%, recall of 93.10%, F1 score of 92.78%, and achieved minimum bias detection time of 16.9ms. Thus, the proposed ADCGN model demonstrates strong performance in mitigating bias within financial chatbots, contributing to more equitable automated financial interactions. This research reinforces the integration of graph-based deep learning within computational economics and supports the development of reliable financial decision support systems.</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

Bias Detection and Mitigation in Financial Chatbots Using Adaptive Deep Crested Porcupine Graph Neural Network

  • 期刊论文

  • 发表日期：

    2026年1月3日

查看全文
  • Veera Prathap Reddy Mamidi,
  • K. S. Sudeep

摘要

Approaches constructed using GPT have enormous promise in financial applications in the time of rapidly developing NLP. However, their integration with financial datasets is still at odds, especially when it comes to evaluating the effectiveness and applicability of such models. To handle these problems, this paper proposes a novel bias detection and mitigation framework powered by an Adaptive Deep Crested porcupine Graph neural Network (ADCGN). It leverages graph neural networks’ powerful ability to learn complex relations in multimodal data-textual and numerical inputs-to extensive bias analysis. It incorporates three major stages: pre-processing through BERT-based MobileNetV3 for cleaning the data and extracting relevant features, in-processing through the proposed module for detecting bias, and postprocessing to mitigate the biases in large language models. It was then experimentally verified using FinGPT-a financial chatbot-showing very superior performance across fairness metrics, reducing bias with improvements in accuracy of 94.56%, precision of 92.47%, recall of 93.10%, F1 score of 92.78%, and achieved minimum bias detection time of 16.9ms. Thus, the proposed ADCGN model demonstrates strong performance in mitigating bias within financial chatbots, contributing to more equitable automated financial interactions. This research reinforces the integration of graph-based deep learning within computational economics and supports the development of reliable financial decision support systems.