Scholarlink

<p>Automated generation and validation of UML diagrams from natural language requirements remain a significant challenge in software engineering, hindered by semantic ambiguity and the lack of robust evaluation methods. This paper introduces a novel, end-to-end pipeline that automates UML modeling by synergistically integrating three specialized AI model types. Our pipeline first employs an instruction-tuned model (LLaMA−3.2-1B) to generate user-centric feature descriptions. These are then processed by a reasoning-centric LLM (DeepSeek-R1-Distill-Qwen-32B) to synthesize PlantUML code through structured reasoning. Finally, an ensemble of Vision-Language Models (VLMs) performs a weighted, multimodal evaluation to ensure the semantic and structural correctness of the generated diagrams. Experiments on a large-scale dataset of 15,000 diagrams across nine types show our pipeline achieves a high Mean Weighted Score of 4.23 on a 1–6 VLM evaluation scale. A human evaluation with 155 participants confirms a strong correlation between our automated scores and expert judgment. Our work establishes a new benchmark for generalized UML automation and demonstrates the effectiveness of a multi-stage pipeline that combines instruction tuning, reasoning-based code synthesis, and multimodal validation.</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

A novel automated method for UML modeling integrating instruction, reasoning LLMs and vision language model evaluation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月26日

查看全文
  • Van-Viet Nguyen,
  • Huu-Khanh Nguyen,
  • Kim-Son Nguyen,
  • Thi Minh-Hue Luong,
  • Anh-Tu Bui,
  • Duc-Quang Vu,
  • The-Vinh Nguyen

摘要

Automated generation and validation of UML diagrams from natural language requirements remain a significant challenge in software engineering, hindered by semantic ambiguity and the lack of robust evaluation methods. This paper introduces a novel, end-to-end pipeline that automates UML modeling by synergistically integrating three specialized AI model types. Our pipeline first employs an instruction-tuned model (LLaMA−3.2-1B) to generate user-centric feature descriptions. These are then processed by a reasoning-centric LLM (DeepSeek-R1-Distill-Qwen-32B) to synthesize PlantUML code through structured reasoning. Finally, an ensemble of Vision-Language Models (VLMs) performs a weighted, multimodal evaluation to ensure the semantic and structural correctness of the generated diagrams. Experiments on a large-scale dataset of 15,000 diagrams across nine types show our pipeline achieves a high Mean Weighted Score of 4.23 on a 1–6 VLM evaluation scale. A human evaluation with 155 participants confirms a strong correlation between our automated scores and expert judgment. Our work establishes a new benchmark for generalized UML automation and demonstrates the effectiveness of a multi-stage pipeline that combines instruction tuning, reasoning-based code synthesis, and multimodal validation.