dahutapea/Topic-Modeling-Tokopedia-Reviews

Tokopedia Review Insights

NLP topic modeling on Indonesian Tokopedia product reviews (PRDECT-ID). The app surfaces what customers praise and complain about - not just sentiment, but the theme behind it.

Pages:

• —EDA - sentiment balance, themes, emotions, categories.
• —Predict Theme - paste a review, get its predicted theme (praise vs complaint).
• —Submit a Review - save a new review to the database (Supabase).

How prediction works

A review is embedded with paraphrase-multilingual-MiniLM-L12-v2 and cosine-matched to the exported topic vectors (artifacts/topic_index.npz). No BERTopic/UMAP/HDBSCAN are needed at runtime, so the Space stays light.

Run locally

pip install -r requirements.txt
streamlit run src/app.py

Deploy to Hugging Face Spaces

1. 1.Create a new Space -> SDK: Streamlit.
2. 2.Upload these files/folders (NOT the notebook, raw dataset, or BERTopic models):
3. 3.src/
4. 4.artifacts/topic_index.npz, artifacts/topic_index.json, artifacts/dashboard_data.csv
5. 5.colloquial-indonesian-lexicon.csv
6. 6.requirements.txt, README.md
7. 7.Set the database credentials under Space -> Settings -> Secrets (do NOT commit secrets.toml). Add a secret named secrets.toml-style connection, or set the [connections.supabase] values via the Secrets UI.

Security

.streamlit/secrets.toml holds DB credentials and is gitignored. Use .streamlit/secrets.toml.example as a template. Never commit real credentials to a public repo.

Project

Hacktiv8 FTDS-040-HCK - Final Project - Group 001.