Electronic Theses and Dissertation

Ulasan produk pada platform e-commerce Shopee berjumlah sangat besar sehingga calon pembeli kesulitan membaca keseluruhan ulasan untuk mengambil keputusan pembelian. Penelitian ini bertujuan mengimplementasikan model Bidirectional and Auto-Regressive Transformers (BART) sebagai sistem yang merangkum banyak ulasan produk berbahasa Indonesia menjadi satu ringkasan abstraktif yang singkat. Dataset dikumpulkan dengan teknik scraping melalui Application Programming Interface (API) Shopee sebanyak 1.048.575 ulasan dari 621 produk pada 21 kategori, kemudian melalui tahap prapemrosesan berupa pembersihan teks, normalisasi kata tidak baku, dan penggabungan ulasan bertingkat per rating hingga menghasilkan 2.673 baris data siap latih. Setiap baris data diberi satu ringkasan acuan yang dibentuk melalui distilasi pengetahuan berbasis Large Language Model (LLM). Tiga model kandidat, yaitu mBART-50 sebagai perwakilan keluarga BART serta cahya-t5 dan IndoNanoT5-base sebagai pembanding berbasis T5, dilatih menggunakan metode fine-tuning hemat parameter Low-Rank Adaptation (LoRA), lalu dievaluasi menggunakan metrik Recall-Oriented Understudy for Gisting Evaluation (ROUGE) dan Bilingual Evaluation Understudy (BLEU) pada data uji. Hasil evaluasi menunjukkan model mBART-50 mencapai performa terbaik dengan ROUGE-1 sebesar 73,28, ROUGE-2 sebesar 50,24, ROUGE-L sebesar 63,86, dan BLEU sebesar 43,99, unggul atas kedua model pembanding pada seluruh metrik. Metode LoRA terbukti mampu mengurangi penggunaan memori GPU selama pelatihan karena hanya melatih 0,77% hingga 5,48% parameter model. Model terbaik selanjutnya diimplementasikan ke dalam prototipe aplikasi web lokal berbasis React dan FastAPI yang mampu menghasilkan ringkasan dalam 12 hingga 45 detik pada perangkat tanpa GPU. Temuan ini menunjukkan bahwa model BART yang menjalani proses fine-tuning dengan metode LoRA efektif dalam meringkas ulasan produk Shopee berbahasa Indonesia.

Product reviews on the Shopee e-commerce platform are extremely numerous, making it difficult for prospective buyers to read all of the reviews before making a purchase decision. This study aims to implement the Bidirectional and Auto-Regressive Transformers (BART) model as a system that condenses a large number of Indonesian-language product reviews into a single concise abstractive summary. The dataset was collected through Shopee Application Programming Interface (API) scraping, comprising 1,048,575 reviews from 621 products across 21 categories, which then went through preprocessing stages consisting of text cleaning, slang normalization, and rating-stratified review bundling, resulting in 2,673 training-ready data rows. Each data row was paired with one reference summary produced through Large Language Model (LLM) based knowledge distillation. Three candidate models, namely mBART-50 as the representative of the BART family, along with cahya-t5 and IndoNanoT5-base as T5-based baselines, were trained using the parameter-efficient fine-tuning method Low-Rank Adaptation (LoRA), and then evaluated using the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) and Bilingual Evaluation Understudy (BLEU) metrics on the test set. The evaluation results show that the mBART-50 model achieved the best performance with a ROUGE-1 of 73.28, ROUGE-2 of 50.24, ROUGE-L of 63.86, and BLEU of 43.99, outperforming both baseline models on every metric. The LoRA method was shown to reduce GPU memory usage during training, as it trains only 0.77% to 5.48% of the model parameters. The best model was then implemented into a local web application prototype based on React and FastAPI, capable of generating a summary within 12 to 45 seconds on a device without a GPU. These findings indicate that the BART model, fine-tuned using the LoRA method, is effective in summarizing Indonesian language Shopee product reviews.