Rancang Bangun Aplikasi Pembangkitan Laporan ESG Menggunakan LLM

Athallarik, Haikal (2026) Rancang Bangun Aplikasi Pembangkitan Laporan ESG Menggunakan LLM. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221232_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Pelaporan Environmental, Social, and Governance (ESG) telah menjadi instrumen fundamental bagi perusahaan untuk menjamin transparansi keberlanjutan dan memenuhi standar regulasi global. Namun proses evaluasi kepatuhannya saat ini masih sangat bergantung pada audit manual yang mahal, memakan waktu, dan rentan terhadap inkonsistensi penilaian akibat subjektivitas manusia. Untuk mengatasi permasalahan tersebut, dikembangkan sistem otomatisasi penilaian ESG berbasis Large Language Model (LLM) yang menerapkan Sequential Prompting dengan tiga tahapan berurutan, yaitu analisis kesenjangan, penilaian skor, dan pembangkitan narasi per kategori ESG. Sistem dibangun menggunakan backend Go dan frontend Next.js, dengan teknik rekayasa prompt yang berbeda untuk setiap tahapan, meliputi pencocokan semantik berbasis kumpulan bukti untuk analisis kesenjangan, Chain-of-Verification (CoVe) untuk penilaian skor, dan Financial Chain-of-Thought (FinCoT) untuk pembangkitan narasi. Evaluasi terhadap 71 indikator ESG dari dokumen terstruktur menunjukkan akurasi analisis kesenjangan sebesar 29,6% dengan makro F1-Score 0,290, Mean Absolute Error penilaian skor sebesar 51,3 poin dengan korelasi Spearman sebesar 0,188, dan rata-rata kualitas narasi sebesar 2,47 dari evaluator non-ahli dan 2,80 dari evaluator ahli. Keterbatasan akurasi bersumber dari perbedaan antara dokumen publik yang digunakan sistem dengan laporan implementasi internal perusahaan, bukan dari kemampuan penalaran model.
===============================================================================================================================
Environmental, Social, and Governance (ESG) reporting has become a fundamental tool for companies to ensure sustainability transparency and meet global regulatory standards. However, the process of evaluating compliance is currently still heavily reliant on manual audits, which are costly, time-consuming, and prone to inconsistencies in assessment due to human subjectivity. To address this problem, an LLM-based ESG assessment automation system was developed, implementing a Sequential Prompting architecture across three ordered stages, which is gap analysis, score assessment, and narrative generation per ESG category. The system was built using a Go backend and Next.js frontend, applying distinct prompt engineering techniques for each stage, semantic matching based on evidence pools for gap analysis, Chain-of-Verification (CoVe) for score assessment, and Financial Chain-of-Thought (FinCoT) for narrative synthesis. Evaluation against 71 ESG indicators from structured documents yielded a gap analysis accuracy of 29.6% with a macro F1-Score of 0.290, a score assessment Mean Absolute Error of 51.3 points and Spearman correlation of 0.188, and an average narrative quality score of 2.47 from the novice evaluator and 2.80 from the expert evaluator. The observed accuracy limitations are primarily attributed to differences between the public documents accessible to the system and the company's internal implementation records, rather than limitations in the model's reasoning capability.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Large Language Model (LLM), Penilaian ESG, Sequential Prompting, Financial Chain-of-Thought (FinCoT), Chain-of-Verification (CoVe), Analisis Kesenjangan, ESG Assessment, Gap Analysis
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.754 Software architecture. Computer software T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK5105.888 Web sites--Design. Web site development.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Haikal Athallarik
Date Deposited:	28 Jul 2026 01:07
Last Modified:	28 Jul 2026 01:07
URI:	http://repository.its.ac.id/id/eprint/137899

Actions (login required)

View Item