Pemanfaatan Large Language Model untuk Pembangkitan Model Domain berdasarkan Paradigma Domain-Driven Design

Syamsiyah, Ainun Nadhifah (2026) Pemanfaatan Large Language Model untuk Pembangkitan Model Domain berdasarkan Paradigma Domain-Driven Design. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221053-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB)

Abstract

Pemodelan domain bertujuan merepresentasikan konsep, hubungan, dan aturan bisnis sebagai dasar pengembangan perangkat lunak. Namun, transformasi deskripsi kebutuhan menjadi model domain memerlukan waktu yang lama, rentan kesalahan, dan membutuhkan pemahaman yang mendalam. Domain-Driven Design (DDD) menawarkan pendekatan untuk membangun rich domain model yang merepresentasikan logika dan aturan bisnis secara eksplisit. Meskipun penelitian sebelumnya menunjukkan bahwa Large Language Model (LLM) dapat mengotomatisasi domain modeling, keluaran yang dihasilkan masih berupa representasi tekstual atau struktur statis yang belum mengimplementasikan prinsip DDD secara lengkap. Tugas akhir ini mengembangkan pendekatan otomatisasi pembangkitan model domain berbentuk code snippets dalam bahasa pemrograman C# menggunakan LLM dengan teknik prompt engineering melalui arsitektur Multi-Step Iterative Generator (MIG) yang terdiri dari 8 langkah Domain Model Generation dan 8 langkah Code Snippets Generation. Evaluasi dilakukan terhadap tiga skenario konfigurasi LLM, yaitu konfigurasi general-purpose, coding-purpose, dan general-coding, menggunakan metrik Test Pass Rate (TPR) untuk mengukur kebenaran fungsional, dilengkapi analisis tipe kegagalan dua tingkat secara kualitatif, yaitu tingkat kompilasi dan tingkat semantik, pada 8 dataset deskripsi kebutuhan. Hasil evaluasi menunjukkan bahwa konfigurasi general-coding yang memadukan model general-purpose (GPT-5.2) pada tahap Domain Model Generation dan model coding-purpose (Qwen 3.5 397B A17B) pada tahap Code Snippets Generation mencapai rata-rata TPR tertinggi sebesar 0,978 dengan jumlah kegagalan unit test paling sedikit, lebih tinggi dibandingkan konfigurasi general-purpose (0,972) dan coding-purpose (0,946). Temuan ini menunjukkan bahwa pemisahan peran model sesuai karakteristik setiap tahap menghasilkan performa lebih baik dibandingkan penggunaan satu jenis model pada seluruh tahapan. Pada tingkat kompilasi, keluaran mentah konfigurasi yang menggunakan model coding-purpose pada tahap Code Snippets Generation lebih sering memerlukan perbaikan, yaitu gagal terkompilasi pada tujuh dari delapan spesifikasi dibandingkan tiga dari delapan pada model general-purpose, sedangkan pada tingkat semantik Business Rule Not Enforced merupakan kegagalan paling dominan dan terkonsentrasi pada konfigurasi coding-purpose, karena penggunaan model coding-purpose pada tahap Domain Model Generation kurang konsisten menegakkan aturan bisnis dan kesalahan tersebut terbawa ke tahap pembangkitan kode. Selain itu, tugas akhir ini membangun 8 dataset text-to-code berdasarkan teks deskripsi kebutuhan untuk keperluan pembangkitan model domain sebagai kontribusi bagi penelitian selanjutnya.
======================================================================================================================================
Domain modeling aims to represent the concepts, relationships, and business rules of a problem domain as the foundation for software development. However, transforming requirement descriptions into domain models remains time-consuming, error-prone, and requires substantial domain expertise. Domain-Driven Design (DDD) provides an approach for constructing rich domain models that explicitly capture business logic and rules. Although previous studies have shown that Large Language Models (LLMs) can automate domain modeling, their outputs are generally limited to textual representations or static structures that do not fully implement DDD principles. This final project develops an automated approach for generating domain models as C# code snippets using LLMs with prompt engineering techniques through a Multi-Step Iterative Generator (MIG) architecture consisting of 8 steps for Domain Model Generation and 8 steps for Code Snippets Generation. Evaluation was conducted on three LLM configuration scenarios, namely general-purpose, coding-purpose, and general-coding, using the Test Pass Rate (TPR) metric to measure functional correctness, complemented by a two-level qualitative failure-type analysis at the compilation and semantic levels across 8 requirement specification datasets. Results show that the general-coding configuration that combining a general-purpose model (GPT-5.2) for Domain Model Generation and a coding-purpose model (Qwen 3.5 397B A17B) for Code Snippets Generation achieves the highest average TPR of 0.978 with the fewest failing unit tests, outperforming the general-purpose (0.972) and coding-purpose (0.946) configurations. These findings indicate that separating model roles according to the characteristics of each stage yields better performance than using a single model type across all stages. At the compilation level, raw output from configurations using the coding model for Code Snippets Generation more often required fixes, failing to compile on seven of eight specifications versus three of eight for the general-purpose model, whereas at the semantic level Business Rule Not Enforced is the most dominant failure and is concentrated in the coding-purpose configuration, because using the coding model for Domain Model Generation less consistently enforces business rules and these errors propagate into the code-generation stage. Additionally, this final project constructs 8 text-to-code datasets from requirement description texts for domain model generation, as a contribution for future research.

Item Type:	Thesis (Other)
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.754 Software architecture. Computer software Q Science > QA Mathematics > QA76.758 Software engineering
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Ainun Nadhifah Syamsiyah
Date Deposited:	27 Jul 2026 02:46
Last Modified:	27 Jul 2026 02:46
URI:	http://repository.its.ac.id/id/eprint/137680

Actions (login required)

View Item