Article Per Year (5 Year)
p-Index From 2021 - 2026
0.23
P-Index
This Author published in this journals
All Journal RABIT: Jurnal Teknologi dan Sistem Informasi Univrab
Niko Felix Chandra Arisco
Universitas Mercu Buana Yogyakarta
Computer Science & IT Engineering

Published : 1 Documents Claim Missing Document
Claim Missing Document
Check
Articles

Found 1 Documents
Search

 1 
EKSTRAKSI INFORMASI NOTA BELANJA INDONESIA MENGGUNAKAN INDOBERT DAN PENANGANAN NOISE BERBASIS FUZZY STRING MATCHING Niko Felix Chandra Arisco; Agus Sidiq Purnomo
Rabit : Jurnal Teknologi dan Sistem Informasi Univrab Vol 11 No 2 (2026): Juli
Publisher : LPPM Universitas Abdurrab

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.36341/rabit.v11i2.8103

Abstract

Indonesian shopping receipts have essential information like items, prices, taxes, and totals; however, their language is unstructured and susceptible to Optical Character Recognition (OCR) errors, such as misidentified characters and incomplete tokens. These problems make entity extraction less accurate, especially on numeric fields. Also, rule-based methods don't scale well, and deep learning models need a lot of labelled data. This study concentrates on developing a resilient information extraction system capable of handling text noise in Indonesian receipts. This method refines IndoBERT for Named Entity Recognition utilising the Consolidated Receipt Dataset (CORD), which has been manually re-annotated using the BIO scheme to identify four entities: ITEM, PRICE, TAX, and TOTAL, comprising 800 training instances, 100 validation instances, and 100 test instances, alongside Levenshtein Distance-based Fuzzy String Matching and rule-based validation. The fuzzy module is evaluated in two configurations, pre-processing and post-processing, via a six-scenario ablation study (S1-S6) in conjunction with OCR typo augmentation. The test findings indicate that the optimal configuration is S6 (IndoBERT + augmentation + FSM post-processing + rule-based validation), achieving a macro F1-Score of 0.7807, surpassing both the pre-processing placement (S5: 0.7733) and the pure IndoBERT baseline (0.7601). Rule-based validation yields the greatest contribution, with an F1 improvement of up to +0.028 due to enhanced recall. The conclusion is that fuzzy string matching is more effective as a post-processing technique, and the pipeline is implemented in a Streamlit-based online prototype that generates structured JSON output.
Co-Authors Sidiq Purnomo, Agus