BPMN Assistant: An LLM-Based Approach to Business Process ModelingBPMN Assistant：一种基于大语言模型的业务流程建模方法

Business Process Model and Notation (BPMN) has long served as a standard for modeling business processes, enabling organizations to visualize and optimize their workflows. However, the complexity inherent in creating, editing, and interpreting BPMN diagrams poses significant challenges, particularly for individuals without specialized training [1]. As Recker [2] demonstrates through extensive research, BPMN’s over-engineered nature and the prevalent lack of formal training among its users create substantial barriers to effective adoption. This bottleneck often results in inefficiencies and a reliance on experts, which can increase operational costs and delay decision-making.业务流程模型和标注（BPMN）长期以来一直是业务流程建模的标准，使组织能够可视化并优化其工作流。然而，创建、编辑和解读 BPMN 图所固有的复杂性带来了巨大挑战，尤其是对于缺乏专业培训的个人而言 [1]。正如 Recker [2] 通过大量研究所证明的那样，BPMN 的过度工程化性质及其用户普遍缺乏正规培训，构成了有效应用该标准的重大障碍。这种瓶颈往往导致效率低下并产生对专家的依赖，进而增加运营成本并延误决策。

A significant challenge in modern organizations is the communication gap between IT departments and business stakeholders. While IT professionals are comfortable with formal modeling notations and technical specifications, business users typically express processes in natural language and informal descriptions. A systematic review by [3] demonstrates that this disconnect leads to misunderstandings, requirements misalignment, and implementation delays, with communication barriers persisting despite significant investments in alignment initiatives. Furthermore, valuable process knowledge frequently remains trapped in the minds of domain experts or scattered across various informal documents, making it difficult to capture and formalize this expertise [4].现代组织面临的一个重大挑战是 IT 部门与业务利益相关者之间的沟通鸿沟。IT 专业人员习惯于使用正式的建模标注和技术规范，而业务用户通常以自然语言和非正式描述来表达流程。[3] 的一项系统性综述表明，这种脱节会导致误解、需求不匹配和实施延误，尽管组织在协同倡议上投入巨大，但沟通障碍依然存在。此外，宝贵的流程知识往往留存在领域专家的脑海中，或分散在各种非正式文档中，使得捕获和形式化这些专业知识变得困难 [4]。

The complexity of extracting process models from unstructured text remains a significant challenge, as highlighted by Bellan et al.’s qualitative analysis [5]. Their research reveals that current methodologies primarily rely on ad-hoc rule-based approaches, which struggle with the complexity of real-world documents. This challenge is further compounded by the limitations of traditional process elicitation methods, as demonstrated by Baião et al. [6], who propose a novel approach integrating group storytelling techniques with text mining to capture the nuances of human-centric activities in process modeling.从非结构化文本中提取流程模型的复杂性仍然是一项重大挑战，正如 Bellan 等人的定性分析所强调的那样 [5]。他们的研究揭示，当前的方法主要依赖于临时性的基于规则的方法，这些方法难以应对现实世界文档的复杂性。正如 Baião 等人 [6] 所展示的那样，传统流程获取方法的局限性进一步加剧了这一挑战，他们提出了一种将群体叙事技术与文本挖掘相结合的新颖方法，以捕获流程建模中以人为中心的活动细微差别。

Recent research has highlighted the cognitive challenges faced by modelers when creating process models, particularly regarding cognitive load and task complexity. Weber et al. [7] introduced a psycho-physiological approach to assess cognitive load during process modeling, utilizing real-time eye movement analysis to identify task-specific difficulties. Their findings emphasize that process modeling requires substantial cognitive effort, especially in naming activities and managing complex structures. These insights reinforce the need for tools like BPMN Assistant, which aim to reduce cognitive barriers and enhance accessibility.近期的研究强调了建模者在创建流程模型时所面临的认知挑战，特别是在认知负荷和任务复杂度方面。Weber 等人 [7] 引入了一种心理生理学方法来评估流程建模过程中的认知负荷，利用实时眼动分析来识别特定任务的困难。他们的研究结果强调，流程建模需要大量的认知投入，特别是在命名活动和管理复杂结构方面。这些见解强化了对 BPMN Assistant 这类工具的需求，旨在降低认知门槛并提高易用性。

The technical barrier to BPMN usage, when combined with the dispersed nature of process documentation, creates a significant barrier to effective process management. Organizations often find themselves in a situation where valuable process knowledge exists but remains underutilized due to its informal and distributed nature.BPMN 使用的技术门槛，加上流程文档的分散性，构成了有效流程管理的重大障碍。组织往往发现自己处于这样一种境地：宝贵的流程知识虽然存在，但由于其非正式和分散的性质而未得到充分利用。

Another critical issue is the dynamic nature of business processes in today’s rapidly evolving business environment. Organizations need to frequently update and modify their processes to remain competitive and adapt to changing market conditions [8]. However, the formal nature of BPMN and the expertise required to modify process models often create a bottleneck in implementing these changes, leading to a gap between actual business operations and their formal documentation. This challenge is evidenced by Leopold et al.’s analysis of 585 BPMN models from industry, which revealed persistent quality issues related to model complexity, particularly in areas such as splits and joins, message flows, and model decomposition, highlighting the technical expertise required for effective BPMN modeling [9].另一个关键问题是当今快速发展的商业环境中业务流程的动态性。组织需要频繁更新和修改其流程以保持竞争力并适应不断变化的市场条件 [8]。然而，BPMN 的形式化性质以及修改流程模型所需的专业知识往往成为实施这些变革的瓶颈，导致实际业务运营与其正式文档之间产生差距。Leopold 等人对工业界 585 个 BPMN 模型的分析证明了这一挑战，该分析揭示了与模型复杂性相关的持续性质量问题，特别是在分支与合并、消息流和模型分解等领域，突显了有效进行 BPMN 建模所需的技术专长 [9]。

The disconnect between formal process models and actual business operations is further exacerbated by the increasing complexity of modern business processes. These processes often span multiple departments, involve numerous stakeholders, and integrate with various systems and external partners. As van der Aalst and Weijters [10] demonstrate in their seminal work on process mining, organizations face significant challenges in accurately capturing and analyzing these complex processes, including issues with hidden tasks, duplicate activities, and non-free-choice constructs. The traditional modeling approaches struggle to address these challenges while maintaining accessibility for all stakeholders.现代业务流程日益增加的复杂性进一步加剧了正式流程模型与实际业务运营之间的脱节。这些流程通常跨越多个部门，涉及众多利益相关者，并与各种系统和外部合作伙伴集成。正如 van der Aalst 和 Weijters [10] 在他们关于流程挖掘的开创性著作中所展示的那样，组织在准确捕获和分析这些复杂流程时面临重大挑战，包括隐藏任务、重复活动和非自由选择结构等问题。传统的建模方法在试图解决这些挑战的同时，难以保持对所有利益相关者的易用性。

Prior to the emergence of Large Language Models (LLMs), numerous attempts were made to automate aspects of process modeling and bridge the natural language gap. Early work by Friedrich et al. [11] demonstrated the potential of rule-based approaches for extracting process models from natural language text, though these systems struggled with complex sentence structures and domain-specific terminology. The integration of natural language processing into process management faced significant challenges with semantic ambiguity and contextual understanding. Efforts by Leopold et al. [12] to automatically generate natural language descriptions from process models highlighted both the potential and limitations of traditional NLP approaches in process automation. These early automation attempts, while groundbreaking, were often constrained by their reliance on predefined rules and limited ability to handle variations in natural language expressions [13].在大语言模型（LLM）出现之前，人们曾多次尝试将流程建模的某些方面自动化，以弥合自然语言带来的鸿沟。Friedrich 等人 [11] 的早期工作展示了基于规则的方法从自然语言文本中提取流程模型的潜力，尽管这些系统在处理复杂句式和领域特定术语时显得吃力。将自然语言处理集成到流程管理中，面临着语义歧义和上下文理解方面的重大挑战。Leopold 等人 [12] 尝试从流程模型自动生成自然语言描述，这既突显了传统 NLP 方法在流程自动化方面的潜力，也暴露了其局限性。这些早期的自动化尝试虽然具有开创性，但往往受限于对预定义规则的依赖，且处理自然语言表达多样性的能力有限 [13]。

Advancements in artificial intelligence, particularly the emergence of LLMs, have opened new avenues for automating the creation and management of business process models [14, 15, 16]. These models have demonstrated exceptional capabilities in understanding and generating natural language, making them well-suited for bridging the gap between textual process descriptions and formal BPMN representations. Recent research by Klievtsova et al. [17] has shown that AI-assisted approaches to process modeling can be particularly effective, with AI-generated models often being preferred over those created by inexperienced human modelers. However, Rebmann et al. [18] highlight that while LLMs show promise in process-related tasks, their effectiveness often depends on proper fine-tuning and task-specific training, particularly for complex semantics-aware operations. This suggests significant potential for LLM-based tools in democratizing access to process modeling capabilities, while also emphasizing the importance of careful implementation and training approaches.人工智能的进步，特别是大语言模型（LLM）的出现，为自动化创建和管理业务流程模型开辟了新途径 [14, 15, 16]。这些模型在理解和生成自然语言方面表现出卓越的能力，非常适合弥合文本流程描述与正式 BPMN 表示之间的鸿沟。Klievtsova 等人 [17] 的近期研究表明，人工智能辅助的流程建模方法非常有效，人工智能生成的模型往往比缺乏经验的人类建模者创建的模型更受欢迎。然而，Rebmann 等人 [18] 强调，虽然 LLM 在流程相关任务中表现出前景，但其有效性往往取决于适当的微调和特定任务的训练，特别是在处理复杂的语义感知操作时。这表明基于 LLM 的工具在普及流程建模能力方面具有巨大潜力，同时也强调了谨慎实施和训练方法的重要性。

This work presents BPMN Assistant, a system that leverages Large Language Models to address the fundamental challenges in business process modeling. While recent research has demonstrated the potential of AI-assisted process modeling, the critical aspect of model modification and maintenance remains largely unexplored. Our work advances the state of the art by demonstrating the effectiveness of LLMs in editing existing BPMN diagrams through natural language instructions.本研究提出了 BPMN Assistant，这是一个利用大语言模型来解决业务流程建模中根本性挑战的系统。尽管近期的研究已经证明了人工智能辅助流程建模的潜力，但模型修改和维护这一关键环节在很大程度上仍未得到探索。我们的工作通过证明 LLM 在通过自然语言指令编辑现有 BPMN 图方面的有效性，推动了该领域的发展。

Our evaluation reveals significant improvements in editing success rates compared to baseline approaches, particularly when using structured representations for model manipulation. These results suggest that LLMs can effectively bridge the gap between informal process descriptions and formal models, offering a promising solution for maintaining and updating business process documentation.我们的评估显示，与基准方法相比，编辑成功率有了显著提高，特别是在使用结构化表示进行模型操作时。这些结果表明，LLM 可以有效地弥合非正式流程描述与正式模型之间的鸿沟，为维护和更新业务流程文档提供了一种有前景的解决方案。

The following sections detail our methodology, system implementation, and evaluation results, providing a foundation for future research in AI-assisted process modeling.接下来的章节将详细介绍我们的方法论、系统实现和评估结果，为人工智能辅助流程建模的未来研究奠定基础。

The integration of Large Language Models (LLMs) into Business Process Management (BPM) has led to significant developments in process automation and modeling. The introduction of specialized tools like the BPMN-Chatbot [19] and ProMoAI [20] has demonstrated the practical applications of this technology. Initial explorations by Berti and Qafari [21] into using GPT-4 and Bard (Gemini) for process mining tasks showed promising results in interpreting both procedural and declarative models, highlighting the potential of LLMs in the BPM domain. These developments have led to comprehensive frameworks for process modeling with LLMs [22], showing promising results in both automation and model quality.将大语言模型（LLM）集成到业务流程管理（BPM）中，推动了流程自动化和建模的重大发展。BPMN-Chatbot [19] 和 ProMoAI [20] 等专用工具的推出，展示了该技术的实际应用。Berti 和 Qafari [21] 对使用 GPT-4 和 Bard (Gemini) 进行流程挖掘任务的初步探索显示了令人期待的结果，证明了它们在解释程序化和声明式模型方面的能力，突显了 LLM 在 BPM 领域的潜力。这些进展促成了基于 LLM 的流程建模综合框架 [22] 的诞生，在自动化和模型质量方面均取得了可观的成果。

In the area of process extraction, various approaches have demonstrated different methodologies and results. Ferreira et al. [23] developed a semi-automatic method for identifying business process elements from natural language texts, achieving 91.92% accuracy in their prototype implementation through carefully defined mapping rules. Taking a different approach, Neuberger et al. [24] introduced a universal prompting strategy that leverages LLMs for process model information extraction, demonstrating performance improvements of up to 8% F1 score over traditional machine learning methods.在流程提取领域，各种方法展示了不同的方法论和结果。Ferreira 等人 [23] 开发了一种从自然语言文本中识别业务流程元素的半自动方法，通过精心定义的映射规则，在其原型实现中达到了 91.92% 的准确率。Neuberger 等人 [24] 采取了不同的方法，引入了一种利用 LLM 进行流程模型信息提取的通用提示策略，与传统的机器学习方法相比，F1 分数提升了高达 8%。

The integration of prompt engineering in BPM has emerged as a promising direction, as discussed by Busch et al. [25]. Their research highlights how prompt engineering can effectively utilize pre-trained language models without extensive fine-tuning, addressing common challenges in process extraction and predictive monitoring. This approach is particularly valuable in scenarios with limited data availability, offering a computationally sustainable solution for BPM applications.正如 Busch 等人 [25] 所讨论的那样，将提示工程（Prompt Engineering）集成到 BPM 中已成为一个有前景的方向。他们的研究强调了提示工程如何能够在无需大量微调的情况下有效利用预训练语言模型，从而解决流程提取和预测性监控中的常见挑战。这种方法在数据可用性有限的情况下尤为宝贵，为 BPM 应用提供了一种计算上可持续的解决方案。

ProMoAI, introduced by Kourani et al. [20], represents a significant development in LLM-driven process modeling. While the tool incorporates prompt engineering and error handling techniques, it currently generates Graphviz graph visualizations rather than process models based on standardized BPMN notation. The tool’s effectiveness was thoroughly evaluated in a subsequent study [26], which demonstrated its capabilities when paired with different LLMs, with Claude 3.5 Sonnet emerging as the best-performing model for generating high-quality process models. Despite using non-standard notation, ProMoAI’s approach demonstrates strong potential for using LLMs to assist in process visualization tasks.Kourani 等人 [20] 引入的 ProMoAI 代表了 LLM 驱动的流程建模的一项重大发展。虽然该工具结合了提示工程和错误处理技术，但它目前生成的是 Graphviz 图可视化，而非基于标准化 BPMN 标注的流程模型。随后的研究 [26] 对该工具的有效性进行了全面评估，展示了其在配合不同 LLM 使用时的能力，其中 Claude 3.5 Sonnet 成为生成高质量流程模型表现最好的模型。尽管使用了非标准标注，ProMoAI 的方法仍展示了使用 LLM 辅助流程可视化任务的强大潜力。

Köpke and Safan [19] introduced the BPMN-Chatbot, a publicly available web-based tool designed for efficient LLM-based process modeling. The BPMN-Chatbot allows users to generate BPMN process models interactively using text or voice input. It achieved notable efficiency gains, reducing token usage by up to 94% compared to alternative tools like ProMoAI while maintaining a correctness rate of 95%, which surpassed the best competitor’s 86%. The authors emphasized the importance of user testing to evaluate the feedback loop’s capabilities, which are critical for interactive process design.Köpke 和 Safan [19] 推出了 BPMN-Chatbot，这是一款公开可用的基于 Web 的工具，专为高效的基于 LLM 的流程建模而设计。BPMN-Chatbot 允许用户通过文本或语音输入交互式地生成 BPMN 流程模型。它实现了显著的效率提升，与 ProMoAI 等替代工具相比，Token 使用量减少了高达 94%，同时保持了 95% 的正确率，超过了最强竞争对手的 86%。作者强调了用户测试对于评估反馈循环能力的重要性，这对于交互式流程设计至关重要。

While BPMN-Chatbot demonstrates notable strengths, particularly in token efficiency and correctness rates, its evaluation primarily focuses on the generation of process models. The editing capabilities, which are essential for iterative refinement and interactive process modeling, have not been rigorously evaluated. Addressing this gap could further validate its utility in real-world scenarios where users frequently modify and adapt process models.虽然 BPMN-Chatbot 展现出显著优势，特别是在 Token 效率和正确率方面，但其评估主要集中在流程模型的生成上。对于迭代优化和交互式流程建模至关重要的编辑功能，尚未经过严格评估。解决这一空白可以进一步验证其在用户频繁修改和调整流程模型的现实场景中的实用性。

In a comprehensive evaluation of process extraction approaches, Bellan et al. [27] analyzed ten state-of-the-art methods for extracting process models from textual descriptions. Their systematic comparison revealed significant variations in performance and methodology among existing tools, with no single approach achieving superior results across all evaluation metrics. This study highlighted the need for standardized evaluation frameworks and emphasized the challenges in automated process extraction that our work aims to address.在对流程提取方法的一项全面评估中，Bellan 等人 [27] 分析了十种从文本描述中提取流程模型的最先进方法。他们的系统比较揭示了现有工具在性能和方法论上的显著差异，没有哪一种方法在所有评估指标上都表现优异。这项研究强调了标准化评估框架的需求，并突显了自动化流程提取中我们旨在解决的挑战。

Kourani et al. [22] proposed a comprehensive framework leveraging LLMs to automate the generation and refinement of process models from textual descriptions. This framework demonstrated its ability to streamline process modeling tasks while maintaining sound and executable model outputs. The superiority of this approach was evident in its comparison with traditional methods, particularly in resolving errors and integrating user feedback effectively. GPT-4 showcased strong performance in generating process models, addressing errors, and adapting to user feedback, whereas Gemini struggled with similar tasks.Kourani 等人 [22] 提出了一个综合框架，利用 LLM 自动化从文本描述中生成和优化流程模型。该框架展示了在保持模型健全且可执行的同时简化流程建模任务的能力。与传统方法相比，这种方法的优越性显而易见，特别是在解决错误和有效整合用户反馈方面。GPT-4 在生成流程模型、纠错和适应用户反馈方面表现强劲，而 Gemini 在类似任务中则表现吃力。

Nivon et al. [28] introduced a novel approach to automating BPMN process generation from textual requirements, addressing the challenges faced by non-expert users in process modeling. Their methodology employs a three-step approach utilizing a fine-tuned GPT-3.5 model: first extracting tasks and ordering constraints from textual descriptions, then constructing an abstract syntax tree (AST) to represent task relationships, and finally converting the AST into a BPMN process. Their Java-based implementation achieved a 78.5% accuracy rate for valid BPMN processes in tests across 200 descriptions, demonstrating the viability of automated BPMN generation while highlighting areas for potential improvement through more advanced language models.Nivon 等人 [28] 引入了一种从文本需求中自动化生成 BPMN 流程的新颖方法，解决了非专家用户在流程建模中面临的挑战。他们的方法采用了一个利用微调后的 GPT-3.5 模型的三步走策略：首先从文本描述中提取任务和顺序约束，然后构建一个抽象语法树（AST）来表示任务关系，最后将 AST 转换为 BPMN 流程。他们在 200 个描述的测试中，基于 Java 的实现达到了 78.5% 的 BPMN 有效流程准确率，证明了自动化 BPMN 生成的可行性，同时也指出了通过更先进的语言模型进行改进的潜力。

These studies collectively illustrate the transformative potential of LLMs in business process modeling. Tools like BPMN-Chatbot, ProMoAI, and Nivon et al.’s automated BPMN generator demonstrate the feasibility of combining generative AI with domain-specific methodologies to achieve higher efficiency, accuracy, and accessibility. Moreover, the benchmarks and frameworks established by researchers provide a foundation for future advancements, emphasizing the importance of careful prompt design, user feedback integration, and domain-specific optimization in realizing the full potential of LLMs in BPM.这些研究共同说明了 LLM 在业务流程建模中的变革潜力。BPMN-Chatbot、ProMoAI 和 Nivon 等人的自动化 BPMN 生成器等工具，证明了将生成式 AI 与领域特定方法相结合以实现更高效率、准确性和易用性的可行性。此外，研究人员建立的基准和框架为未来的进步奠定了基础，强调了仔细的提示设计、用户反馈整合以及领域特定优化在实现 LLM 在 BPM 中全部潜力方面的重要性。

The BPMN Assistant system architecture, as illustrated in Figure 1, is designed to facilitate seamless interaction between users and BPMN models through natural language inputs. The system is composed of three primary components: the Python-based backend, the BPMN layout server, and the Vue.js frontend. Each component plays a critical role in ensuring the system’s functionality, efficiency, and usability.如图 1 所示，BPMN Assistant 系统架构旨在通过自然语言输入促进用户与 BPMN 模型之间的无缝交互。该系统由三个主要组件组成：基于 Python 的后端、BPMN 布局服务器和 Vue.js 前端。每个组件在确保系统的功能性、效率和易用性方面都发挥着关键作用。

The evaluation of BPMN Assistant focuses on assessing its accuracy through Graph Edit Distance (GED) and Relative Graph Edit Distance (RGED), two widely recognized metrics for measuring process model similarity [33, 34, 35]. Traditional GED methods often struggle with gateway semantics and execution probabilities, limiting their effectiveness in BPMN similarity measurements [36]. Schoknecht et al. [37] provide a comprehensive review of process model similarity techniques, emphasizing that hybrid approaches—integrating syntactic, semantic, and behavioral comparisons—tend to produce more reliable results. However, many of these methods require domain-specific adaptations to be effective for BPMN.BPMN Assistant 的评估重点是通过图编辑距离（GED）和相对图编辑距离（RGED）来评估其准确性，这两种指标是衡量流程模型相似度的公认标准 [33, 34, 35]。传统的 GED 方法通常难以处理网关语义和执行概率，限制了它们在 BPMN 相似度测量中的有效性 [36]。Schoknecht 等人 [37] 对流程模型相似度技术进行了全面综述，强调混合方法（整合句法、语义和行为比较）往往能产生更可靠的结果。然而，许多这些方法需要领域特定的适配才能对 BPMN 有效。

Our evaluation methodology deliberately focuses on structural and syntactic assessment through graph similarity measures. This approach was selected based on the research objectives of comparing intermediate representations for BPMN generation and modification. It is important to note that this evaluation does not encompass semantic evaluation approaches such as those based on Petri nets or conformance checking from process mining. Similarly, we do not focus on execution performance through simulation studies, nor do we conduct comprehensive usability evaluations. While these aspects represent valuable dimensions for assessing process modeling tools, they fall outside the scope of the current research, which primarily examines the efficacy of different representation approaches in generating structurally accurate BPMN diagrams from natural language descriptions. The selected evaluation metrics allow for direct comparison between JSON and XML approaches while providing insight into the structural correctness of the generated models.我们的评估方法论特意侧重于通过图相似度测量进行结构和句法评估。选择这种方法是基于比较 BPMN 生成和修改中间表示的研究目标。值得注意的是，此评估不包含基于 Petri 网或流程挖掘一致性检查等语义评估方法。同样，我们不侧重于通过模拟研究进行的执行性能评估，也不进行全面的易用性评估。虽然这些方面代表了评估流程建模工具的有价值维度，但它们超出了当前研究的范畴，本研究主要考察不同表示方法在从自然语言描述生成结构准确的 BPMN 图方面的功效。所选的评估指标允许在 JSON 和 XML 方法之间进行直接比较，同时深入了解生成模型的结构正确性。

Our evaluation demonstrates that the JSON-based representation consistently outperforms direct XML manipulation across multiple metrics. Higher success rates in editing tasks (Table 8) suggest that structured intermediate representations are particularly valuable for complex model modifications. Performance differences between models indicate that model selection significantly impacts system effectiveness, with Claude 3.5 Sonnet and o3-mini showing particularly strong performance.我们的评估表明，基于 JSON 的表示方法在多项指标上均持续优于直接操作 XML。在编辑任务中更高的成功率（表 8）表明，结构化的中间表示对于复杂的模型修改尤为重要。模型间的性能差异显示，模型选择对系统效能有着显著影响，其中 Claude 3.5 Sonnet 和 o3-mini 表现尤为出色。

The inverse relationship between input and output token usage in the JSON versus XML approaches (Table 9) reveals interesting trade-offs that likely favor the JSON approach. Although the JSON approach requires more context in the prompts, it produces more concise output and faster processing times. This combination is particularly advantageous since input tokens typically cost less than output tokens, making the JSON approach potentially more cost-effective despite using more prompt tokens.JSON 方法与 XML 方法在输入和输出 Token 使用量上的反向关系（表 9）揭示了有趣的权衡，且结果很可能更有利于 JSON 方法。尽管 JSON 方法在提示词中需要更多的上下文，但它能生成更简洁的输出并缩短处理时间。这种组合优势明显，因为输入 Token 的成本通常低于输出 Token，这使得 JSON 方法尽管使用了更多的提示词 Token，但在成本上可能更具优势。

Importantly, BPMN Assistant is designed to make business process modeling more accessible to users without a technical background. Instead of requiring knowledge of specialized diagramming tools or technical formats like XML, users can simply describe a process in natural language. The system then automatically generates a professional diagram based on this description. This reduces the need for technical expertise, allowing business professionals to contribute directly to process design and updates, which can lead to more accurate and timely documentation.重要的是，BPMN Assistant 的设计初衷是让没有技术背景的用户也能更轻松地进行业务流程建模。用户无需掌握专业的绘图工具或 XML 等技术格式，只需用自然语言描述流程即可。系统会根据该描述自动生成专业的图表。这降低了对技术专长的需求，使业务专业人员能够直接参与流程的设计与更新，从而获得更准确、更及时的文档。

The system’s ability to handle natural language editing commands has significant implications for process modeling workflows. The high success rates for editing tasks suggest that BPMN Assistant could reduce the technical barriers to process model maintenance, potentially leading to more accurate and up-to-date process documentation in practice.系统处理自然语言编辑指令的能力对流程建模工作流具有重要意义。编辑任务的高成功率表明，BPMN Assistant 可以降低流程模型维护的技术门槛，从而在实践中实现更准确、更及时的流程文档化。

Unlike previous tools such as BPMN-Chatbot [19] and ProMoAI [20], our system emphasizes model modification capabilities in addition to generation. A structured JSON representation offers advantages over both direct XML manipulation and alternative graph representations, such as the Graphviz DOT language ¹¹ 1 https://graphviz.org/doc/info/lang.html, particularly in maintaining BPMN standard compliance and supporting complex editing operations.与 BPMN-Chatbot [19] 和 ProMoAI [20] 等过往工具不同，我们的系统在强调模型生成的同时，也突出了模型修改的能力。结构化的 JSON 表示形式相比直接操作 XML 或其他图表表示形式（如 Graphviz DOT 语言）具有明显优势，尤其是在保持 BPMN 标准合规性和支持复杂编辑操作方面。

Although our approach shows promise, several challenges remain. The dependency on LLM quality affects system reliability, and the current implementation’s limited support for BPMN elements restricts its application in complex scenarios. Additionally, the higher input token requirements of our JSON-based approach may impact costs in high-volume applications.尽管我们的方法前景广阔，但仍面临诸多挑战。对大语言模型（LLM）质量的依赖影响了系统的可靠性，且当前实现对 BPMN 元素的支持有限，限制了其在复杂场景中的应用。此外，基于 JSON 的方法对输入 Token 的高需求可能会影响在大规模应用中的成本。

BPMN Assistant, while demonstrating promising results, has several limitations that should be acknowledged. A key limitation is the support for only a subset of BPMN elements (general tasks, user tasks, service tasks, exclusive gateways, parallel gateways, start events, and end events), which restricts the complexity of processes that can be modeled. However, since the focus of this work is on executable process models rather than collaboration diagrams, the absence of pools and lanes does not significantly impact our primary research objectives.虽然 BPMN Assistant 展现了令人期待的结果，但仍存在一些应予以承认的局限性。一个主要局限是对 BPMN 元素的支持仅限于子集（常规任务、用户任务、服务任务、排他网关、并行网关、开始事件和结束事件），这限制了可建模流程的复杂度。然而，由于本研究的重点在于可执行流程模型而非协作图，因此缺乏池（Pool）和泳道（Lane）并不显著影响我们的主要研究目标。

Due to the lack of semantic evaluation, the results should be considered experimental, as there is no guarantee that the business logic will be correctly integrated into the generated models. This limitation is particularly relevant when assessing the practical applicability of the generated diagrams in real-world scenarios.由于缺乏语义评估，研究结果应被视为实验性的，因为无法保证业务逻辑能被正确集成到生成的模型中。在评估生成图表在现实场景中的实际适用性时，这一局限性尤为关键。

The performance of BPMN Assistant is heavily dependent on the underlying LLM. Our evaluation revealed significant variations in performance across different models, with Claude 3.5 Sonnet and o3-mini showing particularly strong results. Further investigation into the specific capabilities and limitations of different LLMs in the context of process modeling would provide valuable insights for future improvements.BPMN Assistant 的性能在很大程度上取决于底层 LLM。我们的评估显示，不同模型间的性能差异显著，其中 Claude 3.5 Sonnet 和 o3-mini 的表现尤为强劲。深入探究不同 LLM 在流程建模背景下的具体能力与局限性，将为未来的改进提供宝贵的见解。

Human-computer interaction (HCI) studies would be necessary to properly evaluate the usability of the system for non-technical users. While BPMN Assistant aims to lower the barrier to process modeling through natural language interaction, comprehensive user studies would be required to validate this claim.若要恰当评估系统对非技术用户的易用性，人机交互（HCI）研究是必不可少的。尽管 BPMN Assistant 旨在通过自然语言交互降低流程建模的门槛，但仍需进行全面的用户研究来验证这一主张。

Another area of concern is the reliance on clear and unambiguous natural language input. Users might encounter difficulties in achieving this clarity, especially in multilingual or context-specific scenarios where language and terminology nuances can introduce ambiguity.另一个令人关注的领域是对清晰且无歧义的自然语言输入的依赖。用户在实现这种清晰度时可能会遇到困难，特别是在多语言或特定上下文的场景中，语言和术语的细微差别可能会引入歧义。

It is important to note that this work deliberately does not focus on semantic evaluation (e.g., using Petri nets) or conformance checking (e.g., process mining techniques), performance and execution time analysis (e.g., simulation studies), or comprehensive usability evaluation. The approach used in this research was selected considering the research objectives, the method of model generation, and existing evaluation methods cited in the literature. Future work could address these limitations by expanding the range of supported BPMN elements, incorporating semantic evaluation techniques, conducting usability studies, and optimizing the system for specific LLMs based on their strengths in process modeling tasks.需要指出的是，本研究特意未将重点放在语义评估（例如使用 Petri 网）、一致性检查（例如流程挖掘技术）、性能与执行时间分析（例如模拟研究）或全面的易用性评估上。本研究采用的方法是基于研究目标、模型生成方式以及文献中引用的现有评估方法而选定的。未来的工作可以通过扩展所支持的 BPMN 元素范围、结合语义评估技术、开展易用性研究，以及根据不同 LLM 在流程建模任务中的优势进行针对性优化，来解决这些局限性。

This paper presented BPMN Assistant, a novel approach to business process modeling that leverages Large Language Models to bridge the gap between natural language descriptions and formal BPMN representations. Through comprehensive evaluation, we demonstrated that our JSON-based representation for BPMN diagrams outperforms direct XML manipulation across multiple key metrics. In generation tasks, the JSON approach achieved lower latency with an average processing time of 14.41 seconds compared to 24.07 seconds for XML generation, while maintaining comparable model quality with an average similarity score of 0.70.本文介绍了 BPMN Assistant，这是一种利用大语言模型弥合自然语言描述与正式 BPMN 表示之间差距的业务流程建模新方法。通过全面评估，我们证明了基于 JSON 的 BPMN 图表表示方法在多项关键指标上均优于直接操作 XML。在生成任务中，JSON 方法实现了更低的延迟，平均处理时间为 14.41 秒，而 XML 生成则为 24.07 秒，同时保持了相当的模型质量，平均相似度得分为 0.70。

The evaluation results particularly highlight the effectiveness of our JSON-based representation in editing tasks, where it achieved consistently higher success rates across all tested models compared to XML-based editing, with editing operations taking an average of 21.46 seconds compared to 46.98 seconds for XML manipulation. While the JSON approach requires higher input token counts, this is offset by significantly more compact outputs, averaging 364.14 tokens compared to XML’s 2,665.32 tokens. These findings suggest that structured intermediate representations can enhance the reliability and efficiency of LLM-based process modeling tools.评估结果尤为突出了基于 JSON 的表示方法在编辑任务中的有效性。在所有测试模型中，该方法在编辑任务上的成功率均持续高于基于 XML 的编辑，且编辑操作平均耗时 21.46 秒，而 XML 操作则需 46.98 秒。尽管 JSON 方法需要更多的输入 Token，但其输出明显更为紧凑，平均为 364.14 个 Token，而 XML 则为 2,665.32 个 Token。这些发现表明，结构化的中间表示可以提升基于 LLM 的流程建模工具的可靠性和效率。

Despite these achievements, our work also reveals important areas for future research. The current limitations in supporting only a subset of BPMN elements indicate opportunities for expanding the system’s capabilities, particularly for modeling more complex executable processes. Additionally, the varying performance across different LLMs suggests potential benefits from model-specific optimization approaches.尽管取得了这些成就，我们的研究也揭示了未来研究的重要领域。目前对 BPMN 元素支持不足的局限性表明，在扩展系统功能方面仍有空间，特别是在建模更复杂的可执行流程方面。此外，不同 LLM 之间性能的差异表明，采用模型特定的优化方法可能带来潜在益处。

In broader terms, this research represents a significant step toward making business process modeling more accessible to a wider audience. By enabling natural language interaction with process modeling tools, BPMN Assistant reduces the technical knowledge required to create and modify business process diagrams. This accessibility has potential implications beyond technical efficiency gains, as it could foster better communication between technical and non-technical stakeholders in organizations. When business experts can directly interact with modeling tools using everyday language, they may become more engaged in the process modeling activities, potentially leading to more accurate process documentation and improved alignment between business needs and implemented processes. The natural language approach also opens possibilities for process modeling in educational contexts, where students can focus on understanding process concepts rather than mastering technical modeling interfaces.从更广泛的角度来看，这项研究代表了向大众普及业务流程建模迈出的重要一步。通过支持流程建模工具的自然语言交互，BPMN Assistant 降低了创建和修改业务流程图所需的技术知识要求。这种易用性带来的影响不仅限于技术效率的提升，还可能促进组织内技术人员与非技术利益相关者之间的沟通。当业务专家能够使用日常语言直接与建模工具交互时，他们可能会更积极地参与流程建模活动，从而可能带来更准确的流程文档，并改善业务需求与已实施流程之间的一致性。自然语言方法也为教育场景中的流程建模开启了可能性，学生可以专注于理解流程概念，而非纠结于掌握复杂的技术建模界面。

Looking ahead, future work could focus on expanding support for additional BPMN elements to enable modeling of more complex executable processes. The significant performance variations observed across different LLMs also suggest opportunities for developing specialized prompt engineering techniques for process modeling tasks. These improvements would further enhance the practical applicability of LLM-based process modeling tools in enterprise settings.展望未来，后续工作可聚焦于扩展对更多 BPMN 元素的支持，以实现对更复杂可执行流程的建模。在不同 LLM 中观察到的显著性能差异也表明，开发针对流程建模任务的专业提示工程技术大有可为。这些改进将进一步增强基于 LLM 的流程建模工具在企业环境中的实际适用性。