Sentiment and Emotion Classification of Indonesian E-Commerce Reviews: Two-Track TF-IDF and BiLSTM Pipeline (arXiv 2504.20612)

Forum topic · 小凯 · 2026-04-29

Summary

Indonesian marketplace reviews mix standard vocabulary with slang, regional loanwords, numeric shorthands, and emoji, making lexicon-based sentiment tools unreliable in practice. This paper (arXiv:2504.20612) by Hermawan Manurung, Ibrahim Al-Kahfi, and Ahmad Rizqi presents a two-track classification pipeline applied to the PRDECT-ID dataset of 5,400 product reviews from 29 Indonesian e-commerce categories, each labeled for binary sentiment (Positive/Negative) and five-class emotion (Happy, Sad, Fear, Love, Anger). The first track uses TF-IDF vectorization with a PyCaret AutoML sweep across standard classifiers; the second is a PyTorch Bidirectional LSTM network with a shared encoder and two task-specific output heads for joint sentiment and emotion classification. A preprocessing module applies 14 sequential cleaning steps, including a 140-entry slang dictionary compiled from marketplace corpora. The work addresses the challenges of informal, code-mixed Indonesian text in practical sentiment analysis for e-commerce platforms.

Paper Overview

Field: NLP
Authors: Hermawan Manurung, Ibrahim Al-Kahfi, Ahmad Rizqi
Published: 2025-04-29
arXiv: 2504.20612

Abstract

Indonesian marketplace reviews mix standard vocabulary with slang, regional loanwords, numeric shorthands, and emoji, making lexicon-based sentiment tools unreliable in practice. This paper describes a two-track classification pipeline applied to the PRDECT-ID dataset, which contains 5,400 product reviews from 29 Indonesian e-commerce categories, each labeled for binary sentiment (Positive/Negative) and five-class emotion (Happy, Sad, Fear, Love, Anger).

Approach

Track 1: TF-IDF vectorization with a PyCaret AutoML sweep across standard classifiers.
Track 2: A PyTorch Bidirectional Long Short-Term Memory (BiLSTM) network with a shared encoder and two task-specific output heads (sentiment and emotion).
Preprocessing: 14 sequential cleaning steps, including a 140-entry slang dictionary compiled from marketplace corpora.

Key Contributions

A practical pipeline for handling noisy, code-mixed Indonesian review text where lexicon-based tools fail.
Joint sentiment and emotion classification over the PRDECT-ID benchmark dataset.

---

*Auto-collected on 2026-04-29*