A Multilevel Classification of Multilingual Negative Reviews in E-commerce
Abstract
A multilingual multi-label framework for identifying complaint-related issue categories from negative e-commerce reviews written in Bangla, English, and Banglish. A 14,280-review corpus was annotated across four issue categories, and eight models spanning classical ML, deep learning, transformers, and hybrid architectures were benchmarked. XLM-RoBERTa achieved the best performance at 97.02% Micro-F1, and the framework was deployed on the IFJONA e-commerce platform.
Research Objectives
- 01Prepare and analyze a multilingual negative review dataset (Bangla, English, Banglish).
- 02Define four complaint categories: Product, Delivery, Monetary, Customer Service.
- 03Formulate complaint detection as a multi-label classification task.
- 04Compare classical ML, deep learning, transformer, and hybrid models.
- 05Evaluate with multi-label metrics — Micro-F1, Macro-F1, Exact Match, Hamming Loss.
- 06Integrate the framework into a real e-commerce review management system.
Dataset at a Glance
Issue Categories
Product Issue
59.17%Quality, damage, wrong item, size/colour mismatch, missing quantity.
Delivery Issue
32.39%Late or failed delivery, damaged parcel, courier mishandling.
Monetary Issue
33.53%Payment failure, delayed refund, overcharging, hidden charges.
Customer Service Issue
32.82%Unresponsive seller, poor support, return/replacement failures.
Model Comparison
Eight models across four architectural families were benchmarked under an identical protocol. XLM-RoBERTa achieved the strongest performance on the held-out test set.
| Rank | Model | Family | Micro-F1 | Macro-F1 |
|---|---|---|---|---|
| 1 | XLM-RoBERTaBest | Transformer | 97.02% | 96.87% |
| 2 | mBERT | Transformer | 96.77% | 96.67% |
| 3 | TF-IDF + Linear SVM | Classical ML | 95.86% | 95.85% |
| 4 | TF-IDF + Logistic Regression | Classical ML | 95.72% | 95.74% |
| 5 | LSTM | Deep Learning | 95.17% | 95.19% |
| 6 | BiLSTM | Deep Learning | 94.67% | 94.52% |
| 7 | BanglaBERT | Transformer | 79.50% | 77.25% |
| 8 | BanglaBERT + BiLSTM | Hybrid | 77.86% | 76.12% |
Real-World Deployment
The framework was integrated into the IFJONA e-commerce platform as a hybrid service-oriented system that classifies incoming negative reviews automatically.
Laravel API
Production backend — review submission, rating filter
FastAPI
ML inference microservice — model loading and prediction
React Admin Panel
Displays predicted issue labels as colour-coded badges
MySQL
Stores reviews and detected issue labels
Thesis Gallery
4 itemsMoments from the thesis defense, the final submission, and the real-time deployment of the model into a live e-commerce admin panel.
Thesis Defense
Presenting the research methodology and results in front of the evaluation panel at Premier University, Chittagong.
With My Team
Post-defense photo with my thesis teammates after successfully completing the final presentation.
Final Thesis Book
The printed and bound copy of the thesis, submitted to the Department of CSE for final approval.
Real-Time Deployment
The thesis model integrated into a live e-commerce admin panel — automatically detecting negative complaint issues from incoming customer reviews in real time.
Publication Status
A manuscript based on this thesis is currently in preparation for journal submission. Peer-reviewed publications will be listed here once accepted.