NLP Data Science 2 Semesters

💬 Malay + Manglish Sentiment Analysis for E-Commerce Reviews

Shopee/Lazada reviews in Malaysia are written in Malay, English, and everything in between ("barang ok tapi delivery lambat gila"). Off-the-shelf sentiment tools are trained on formal English and fail on this. That gap is your research.

1The Pitch — Product + Research in One

🛠️ What you build (the demo)

A web dashboard where a seller pastes or uploads product reviews and gets: sentiment breakdown (positive/negative/neutral), trending complaint keywords, and per-aspect scores (delivery, quality, price, service).

🔬 What you research (the grade)

Which approach handles Malay–English code-switched text best? You benchmark three families of sentiment classifiers on the same datasets and report which wins, where, and why.

2Problem Statement & Objectives

Pernyataan masalah: Malaysian SME sellers receive hundreds of mixed-language reviews they never analyze; existing sentiment tools misclassify code-switched Malay/Manglish text, so sellers can't act on customer feedback.

Build a review-analysis dashboard for Malay/Manglish e-commerce reviews (the MVP).
Benchmark three sentiment approaches — lexicon-based, classical ML, and transformer — on code-switched text.
Evaluate all models with accuracy, precision, recall, and F1 to identify the best approach for Malaysian reviews.

3Benchmark Design — the 3 Experiments

Same shape as a graded FYP: three controlled experiments, one comparison story.

Experiment	Setup	What it proves
Eksperimen 1 Lexicon baseline	VADER + a Malay sentiment lexicon (Malaya) on the review dataset, no training.	Baseline: how far rule-based methods get on code-switched text.
Eksperimen 2 Classical ML	TF-IDF features → Naive Bayes vs SVM vs Logistic Regression, 80/20 train-test split.	Whether lightweight trainable models beat rules when trained on local data.
Eksperimen 3 Transformer	Fine-tuned multilingual BERT (mBERT / MalayBERT) on the same 80/20 split.	Whether pretrained language models justify their cost on Malaysian text.

📦 Dataset plan

Scrape/collect ~2,000–5,000 Shopee/Lazada reviews across 3 product categories; label a subset manually (with a labeling guideline — that's a methodology section by itself). Augment with the public Malaya sentiment corpus for pre-training comparisons — mirroring the "local dataset vs public dataset" comparison pattern examiners love.

4Evaluation Metrics

Metric	Why it matters here
Accuracy	Overall headline number for each of the 3 experiments.
Precision / Recall	Negative-review recall matters most to sellers — missing complaints is costly.
F1-score (macro)	Classes are imbalanced (most reviews are positive) — macro-F1 keeps you honest.
Confusion matrix per language mix	The novel part: slice results by pure-Malay vs pure-English vs code-switched rows.

5Report Mapping (Bab 1–5)

Chapter	What goes in it (from this package)
Bab 1 — Pengenalan	Problem statement above, 3 objectives, scope (3 product categories, 3 model families), Gantt chart.
Bab 2 — Sorotan Kesusasteraan	Comparison tables: sentiment approaches (2019–2026 papers), code-switching NLP studies, Malay NLP resources (Malaya, MalayBERT).
Bab 3 — Metodologi	Data collection + labeling guideline, preprocessing pipeline (emoji, slang normalization), architecture diagram of dashboard + model API.
Bab 4 — Pembangunan & Pengujian	The 3 experiments, metric tables per experiment, per-language-mix analysis, dashboard screenshots.
Bab 5 — Dapatan Kajian	Which approach wins on code-switched text, limitations (labeling subjectivity, dataset size), future work (aspect-based SA).

6Tech Stack

Pythonscikit-learnHugging Face MalayaStreamlit / FlaskPandas

🔎 Related works — Perplexity API placeholder

Live literature search (latest code-switching sentiment papers, 2020–2026) will be generated here via the Perplexity API once the subscription is active.

⏳ Generate Literature Review (coming soon)

← Back to all FYP ideas · Assignment Helper