提出融合大模型与多模态相似度的框架，提升需求重复与冲突检测准确率。
PassionNet: An Innovative Framework for Duplicate and Conflicting Requirements Identification
- 构建三类预测流水线：基于语言模型、多模态相似度驱动、混合式融合。
- 在6个公开数据集上测试760种流水线，混合型模型F1得分领先13%。
- 适合软件工程中需求分析阶段使用，尤其关注质量与效率的团队。
早期发现并解决重复与冲突的需求可显著提升项目效率与整体软件质量。尽管已有研究利用人工智能技术开发多种计算预测方法，但现有方法性能仍不足，亟需更高效的方法支持软件开发流程。为此，本文提出一个综合性框架，支持构建三类预测流水线：基于语言模型、多尺度多模态相似度驱动，以及结合大语言模型上下文与多模态相似度知识的混合型流水线。第一类通过8种不同大语言模型实现需求重复与冲突识别；第二类整合从传统相似度计算到大模型生成的高级相似向量；第三类则融合大模型上下文信息与多模态相似度知识。在6个公共基准数据集上对760种不同预测流水线的广泛测试表明，混合型流水线在准确识别重复与冲突需求方面持续优于其他两类，整体F1分数相比现有最先进方法提升13%。
原文摘要 · Abstract (English)
Early detection and resolution of duplicate and conflicting requirements can significantly enhance project efficiency and overall software quality. Researchers have developed various computational predictors by leveraging Artificial Intelligence (AI) potential to detect duplicate and conflicting requirements. However, these predictors lack in performance and requires more effective approaches to empower software development processes. Following the need of a unique predictor that can accurately identify duplicate and conflicting requirements, this research offers a comprehensive framework that facilitate development of 3 different types of predictive pipelines: language models based, multi-model similarity knowledge-driven and large language models (LLMs) context + multi-model similarity knowledge-driven. Within first type predictive pipelines landscape, framework facilitates conflicting/duplicate requirements identification by leveraging 8 distinct types of LLMs. In second type, framework supports development of predictive pipelines that leverage multi-scale and multi-model similarity knowledge, ranging from traditional similarity computation methods to advanced similarity vectors generated by LLMs. In the third type, the framework synthesizes predictive pipelines by integrating contextual insights from LLMs with multi-model similarity knowledge. Across 6 public benchmark datasets, extensive testing of 760 distinct predictive pipelines demonstrates that hybrid predictive pipelines consistently outperforms other two types predictive pipelines in accurately identifying duplicate and conflicting requirements. This predictive pipeline outperformed existing state-of-the-art predictors performance with an overall performance margin of 13% in terms of F1-score
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。