APPLICATION DEVELOPMENT WITH LLM FOR GENERATION AND VISUALIZATION OF BPMN FROM LEGAL DOCUMENTS

• Main Author: Wilsen, Willy
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86191
• Institution: Institut Teknologi Bandung

Legal documents are often the main foothold in determining the structure and rules governing business processes. The process of creating accurate process models based on complex legal documents requires very careful interpretation and interpretation given the rapid growth of legal documents that are difficult to read and digest all the information. On the other hand, tools such as the Large Language Model (LLM) have brought significant advances in natural language text generation capabilities. These LLMs bring potential in automating the generation of process models from legal documents. The existing research raises the question “How to use pre-trained LLM in applications to generate and visualize legal document workflows that are accurate, easier for users, represent actors clearly, and pay attention to token capacity constraints in processing legal documents?”. The implementation of the solution scheme is based on the draft solution of the problem to provide eight different schemes in the solution architecture. The final solution scheme implemented the use of RAG in the generation of BPMN elements as the workflow of legal documents. The solution architecture using LLM in the form of GPT provides accuracy, precision, and recall rates of 92.85%, 96.42%, and 96.42% respectively and the solution architecture using LLM in the form of Gemini provides accuracy, precision, and recall rates of 67.28%, 81.57%, and 85.71% respectively. However, the exclusiveGateway generated sometimes only gives an if answer. The layout generated by the layout engine to provide the position and size of an element is also still done manually.