declerke/Ecommerce-Review-NLP

NLP intelligence pipeline for Jumia Kenya product reviews - Airflow 3.0 + PostgreSQL + dbt + HuggingFace + Streamlit # 🛍️ E-Commerce Review Sentiment Pipeline: NLP Intelligence for Jumia Kenya **E-Commerce Review Sentiment Pipeline** is a production-grade NLP data engineering solution that scrapes live product reviews from Jumia Kenya across four consumer electronics categories, applies three distinct HuggingFace NLP models to classify overall sentiment, perform four-dimension aspect-based sentiment analysis (ABSA), and score each review for authenticity, then transforms the enriched results through a three-layer dbt pipeline into analytical mart tables surfaced across a four-tab Streamlit dashboard — the complete stack a data science team at an e-commerce platform would use to monitor product reputation, detect suspicious reviews, and understand which product dimensions (price, quality, delivery, service) drive buyer satisfaction. | Metric | Value | |--------|-------| | Reviews scraped | 408 (4 categories · Jumia Kenya) | | Sentiment scored | 408 (cardiffnlp/twitter-roberta-base-sentiment-latest) | | Aspect rows | 1,632 (4 aspects × 408 reviews) | | Authenticity scored | 408 (heuristic scoring) | | Overall positive sentiment | 74.8% | | Avg star rating | 4.10 / 5 | | NLP rating accuracy | 81.1% within 1 star of actual | | Airflow tasks | 6 (scrape → 3× NLP → dbt → cache) | | dbt models | 9 (3 staging · 1 intermediate · 5 marts) | | dbt tests | 44 (all passing) | | Pytest tests | 71 (all passing) | | Dashboard tabs | 4 | | Cost to run | $0 — open-source models + local stack only | --- ## 🎯 Project Goal E-commerce platforms operating in Kenya and across Sub-Saharan Africa collect thousands of product reviews daily, but most of that text data is never analysed — it sits in a database while product managers make decisions based only on star averages. The challenge is not the absence of data but the absence of NLP infrastructure capable of extracting structured signals (sentiment, aspect-level scores, authenticity flags) from unstructured review text at the speed needed for opera …