Bangla and Banglish E-Commerce Reviews Dataset for Aspect-Based Sentiment Analysis

This dataset is a manually annotated collection of 3,587 Bangla and Banglish e-commerce product reviews, curated for Aspect-Based Sentiment Analysis (ABSA) research in low-resource languages. Reviews were collected from Daraz, covering products in the electronics domain such as mobile phones, laptops, earphones, headphones, and CCTV cameras.

The raw dataset initially contained 19,638 reviews, which were preprocessed to remove noise, normalize text, and filter irrelevant entries. Preprocessing steps included removing URLs, emails, and mentions; converting emojis to text; normalizing Banglish; reducing character elongation; ensuring meaningful sentence boundaries; and removing all reviews shorter than 5 words to maintain data quality and reliability. After cleaning, 10,657 reviews were retained.

A subset of 3,587 reviews was manually annotated for ABSA. Each review contains one or more aspect-level sentiment labels (positive, negative, or neutral) covering five aspects: product quality, price, delivery, packaging, and seller service.

Dataset Counts:

Total original reviews: 19,638

Total preprocessed reviews: 10,657

Total annotated reviews: 3,587

Languages (preprocessed):

Bangla: 6,731

Banglish: 2,797

Mixed: 1,129

Folder Structure:

original/ – raw scraped reviews

preprocessed/ – cleaned and normalized reviews, filtered for quality

annotated/ – manually labeled ABSA reviews

metadata/ – summary of review counts, language distribution

Data quality was ensured through duplicate removal, filtering short reviews, manual validation, and annotation consistency checks. This dataset is ready to train and evaluate multilingual transformer models such as BanglaBERT, XLM-RoBERTa, and mBERT for multi-label ABSA tasks, and serves as a benchmark for Bangla–Banglish e-commerce sentiment research.