Vietnamese ABSA — Aspect-Based Sentiment Analysis system for CellphoneS Reviews

An Aspect-Based Sentiment Analysis (ABSA) system for Vietnamese cellphone reviews, covering the full pipeline from data collection and model training to a production-ready web application.

---

Demo

---

System Overview

Crawl data  →  Train models  →  Batch inference  →  Web App
 (Scrapy)      (Notebooks)    (script/inference.py)  (FastAPI + React)

The system is broken into two sub-tasks handled by two fine-tuned models:

Sub-task 1 — Aspect Span Detection (NER)

Identify which text spans mention a specific product aspect.
Model: NlpHUST/ner-vietnamese-electra-base (token classification, BIO tagging)

Aspect	Meaning
BATTERY	Battery life, charging speed
CAMERA	Camera quality, photography features
DESIGN	Appearance, build quality, form factor
FEATURES	Software features, specifications, capabilities
GENERAL	Overall impression, unspecified aspects
PERFORMANCE	Processing speed, lag, heating, stability
PRICE	Price, value for money, promotions
SCREEN	Display quality, brightness, resolution
SER&ACC	Customer service, warranty, accessories
STORAGE	Internal storage, memory capacity

Sub-task 2 — Sentiment Classification (SA)

Classify the sentiment expressed toward each detected aspect (and the review overall).
Model: vinai/phobert-base-v2 (sequence classification, ABSA pair format: <aspect span> : <full text>)

Sentiment	Meaning
POSITIVE	The reviewer expresses satisfaction or praise
NEGATIVE	The reviewer expresses dissatisfaction or criticism
NEUTRAL	The reviewer is neutral or merely descriptive

---

Dataset — UIT-ViSD4SA

Both models are trained on UIT-ViSD4SA, a Vietnamese sentiment dataset for aspect-based analysis of smartphone reviews.

Data format (JSONL):

{ "text": "Man hinh dep, pin trau nhung camera hoi thuong.", "labels": [[0, 8, "SCREEN#POSITIVE"], [10, 19, "BATTERY#POSITIVE"], [28, 34, "CAMERA#NEGATIVE"]] }

---

Model Performance

NER Model — Aspect Span Detection

Fine-tuned from NlpHUST/ner-vietnamese-electra-base. Evaluated with token-level weighted-average F1. Training code: notebook/Electra_NER.ipynb

Aspect	Precision	Recall	F1-score
BATTERY	0.8367	0.8551	0.8458
CAMERA	0.8016	0.8622	0.8308
DESIGN	0.7610	0.8065	0.7831
FEATURES	0.7349	0.7550	0.7448
GENERAL	0.7374	0.7646	0.7508
PERFORMANCE	0.7667	0.7839	0.7752
PRICE	0.5612	0.5901	0.5753
SCREEN	0.6974	0.7567	0.7258
SER&ACC	0.7566	0.7559	0.7563
STORAGE	1.0000	0.0038	0.0077
weighted avg	0.7632	0.7822	0.7700

SA Model — Sentiment Classification

Fine-tuned from vinai/phobert-base-v2. Uses an ABSA pair input format (<aspect span> : <full text>). Training code: notebook/Transformer_SA.ipynb

Sentiment	Precision	Recall	F1-score
NEGATIVE	0.9498	0.9125	0.9307
NEUTRAL	0.5675	0.6944	0.6246
POSITIVE	0.9618	0.9549	0.9583
weighted avg	0.9277	0.9211	0.9238

---

Project Structure

├── backend/
│   ├── main.py          # FastAPI app (REST API + static file serving)
│   └── predictor.py     # NER + SA inference pipeline
├── crawler/
│   ├── spiders/
│   │   └── reviews.py   # Scrapy + Playwright spider for cellphones.com.vn
│   ├── items.py
│   ├── pipelines.py
│   └── settings.py
├── data/
│   ├── crawled/         # Raw reviews from the crawler
│   ├── inferenced/      # ABSA-labeled reviews (output of inference.py)
│   └── train/           # UIT-ViSD4SA splits (train / dev / test JSONL)
├── frontend/            # React 18 + Vite + Tailwind CSS
│   └── src/pages/
│       ├── Dashboard.jsx   # Review browser & analytics dashboard
│       └── Predict.jsx     # Real-time ABSA prediction page
├── models/
│   ├── NER/             # Fine-tuned ELECTRA NER model
│   ├── SA/              # Fine-tuned PhoBERT SA model
│   └── vncorenlp/       # VnCoreNLP word segmenter
├── notebook/
│   ├── Electra_NER.ipynb      # NER model training
│   ├── Transformer_SA.ipynb   # SA model training
│   └── predict.ipynb          # Inference demo
├── script/
│   └── inference.py     # Batch inference over all crawled data
└── pyproject.toml

---

Requirements

• Python >= 3.10
• Node.js >= 18
• GPU with CUDA recommended for faster inference

---

Installation

1. Python dependencies

python -m venv .venv
source .venv/bin/activate       # Windows: .venv\Scripts\activate
pip install -e .

2. Playwright (used by the crawler)

playwright install chromium

3. Frontend dependencies

cd frontend
npm install

---

Usage

Step 1 — Crawl data

cd crawler
scrapy crawl cellphones_reviews -s MAX_PRODUCTS=50 -s MAX_REVIEWS=100

Output saved to data/crawled/cellphones.jsonl.

Parameter	Description	Default
MAX_PRODUCTS	Maximum number of products to crawl	Unlimited
MAX_REVIEWS	Maximum reviews per product	Unlimited

Step 2 — Batch inference

Run the ABSA pipeline over all crawled reviews:

python script/inference.py

Output saved to data/inferenced/cellphones.jsonl.

Step 3 — Run the web app

Start the backend:

uvicorn backend.main:app --reload --port 8000

Start the frontend dev server (in a separate terminal):

cd frontend
npm run dev

Visit http://localhost:5173. API requests are proxied to the backend at http://localhost:8000.

---

API Endpoints

Endpoint	Method	Description
GET /api/products	GET	List all products with review statistics
GET /api/reviews	GET	List reviews with filter & pagination
POST /api/predict	POST	Run ABSA on a given Vietnamese text

Query parameters for /api/reviews:

Parameter	Description
product	Filter by product slug
sentiment	Filter by sentiment (POSITIVE / NEGATIVE / NEUTRAL)
aspect	Filter by aspect (e.g. BATTERY, CAMERA)
page	Page number (default: 1)
size	Items per page, 1–100 (default: 20)

---

Web Interface

• Dashboard — Sentiment distribution statistics, review browser with filtering by product / aspect / sentiment. Aspect spans are highlighted inline within each review.
• Predict — Enter any Vietnamese text and receive real-time ABSA results with color-coded aspect highlights and per-aspect sentiment labels.

---

Tech Stack

Component	Technology
Crawling	Scrapy + Playwright
NLP Preprocessing	VnCoreNLP (word segmentation)
Deep Learning	PyTorch + HuggingFace Transformers
NER Model	ELECTRA (NlpHUST/ner-vietnamese-electra-base)
SA Model	PhoBERT (vinai/phobert-base-v2)
Training Dataset	UIT-ViSD4SA
Backend API	FastAPI + Uvicorn
Frontend	React 18 + Vite + Tailwind CSS
Data format	JSONL