Project Overview

Online reviews play a crucial role in shaping public perception of local businesses. However, the presence of Irrelevant, promotional, or low-quality reviews may distort the true reputation of a place. This project was developed for the TechJam Hackathon challenge on assessing the quality and relevancy of Google location reviews.

We built a classification pipeline that labels reviews into four categories:

• Advertisement (promotions, links, coupons, URLs)
• Irrelevant Content (not related to the business or never visited business)
• Rant (negative opinions to express anger or contempt with little or no constructive feedback)
• Feedback (genuine, relevant review: may be positive/negative)

Our approach combines:

• Zero-shot classification + rule-based overrides for initial labeling
• Feature engineering: TF-IDF (uni/bi-grams) + regex/numeric signals (URLs, promo words, “never been” phrases, punctuation, all-caps ratio, etc.)
• Machine learning models: Logistic Regression + Random Forest
• Evaluation: precision, recall and F1-score

Setup Instructions

Clone the repository

git clone https://github.com/Augeliua/TechJam_ReviewFiltering.git
cd TechJam_ReviewFiltering

Create environment & install dependencies

conda create -n reviews python=3.11 -y
conda activate reviews
pip install -r requirements.txt

Required libraries:

• pandas, numpy, scikit-learn, scipy
• nltk (stopwords, wordnet)
• transformers, torch
• matplotlib, wordcloud
• joblib

Download NLTK resources

import nltk
nltk.download("stopwords")
nltk.download("wordnet")

How to Reproduce Results

1. Prepare the dataset

• Place cleaned_dataset.csv in the project root.
• Ensure it contains:
  • text (review content)
  • label (harmonised into the 4 classes).

2. Run the notebook

• Open notebook.ipynb.
• Execute cells step by step:
  • Imports and downloads
  • Load dataset and preprocess
  • Zero-shot + rule-based overrides
  • Feature extraction (TF-IDF + regex)
  • Model training and evaluation

3. Evaluate models

• Baseline Logistic Regression (TF-IDF 3000) → Accuracy = 91.0%
• Logistic Regression (TF-IDF 30k + regex features) → Accuracy = 95.4%
• Random Forest Baseline → Lower performance than logistic regression however we still chose it as its a more robust model.

4. Make Predictions

import joblib
vec_art = joblib.load("baseline_tfidf3k_logreg.joblib")
vectorizer_base, logreg_base = vec_art["vectorizer_base"], vec_art["logreg_base"]

X = vectorizer_base.transform(["Nice ambience!", "Use code SAVE10 at www.shop.com"])
print(logreg_base.predict(X))

#Team Member Contributions

Name	Contributions
Fong Zhi Lum	Website + frontend + backend
Fong Xin Yi	Website + frontend + backend
Dela Cruz Dionieleth Angelina Carlos\	Data Cleaning + EDA + Notebook Modelling
Chua Ziying	Data Cleaning + EDA + Notebook Modelling
Veda Ho Yong Qian	Data Cleaning + EDA + Notebook Modelling