News Sentiment Analysis using Transformer Models

Research about NLP models performance with news sentimen analysis

Author: Wilbert Winardi (NIM: 2702238716)

📌 Project Overview

This project performs Sentiment Analysis on news articles to classify them into Positive, Neutral, or Negative categories. The primary focus is to compare the efficiency and accuracy of various Transformer-based models to determine the optimal solution for this task.

📂 Dataset Information

• Source: Kaggle - News Sentiment Analysis
• Total Samples: 3,500 rows.
• Features Used: * Input: Description (News summary).
  • Target: Sentiment (Encoded as 0, 1, 2).

🛠️ Methodology

1. Preprocessing: Lowercasing, removing URLs, punctuation ([^\w\d\s\-]), and numbers (\d+) to reduce bias.
2. Vectorization/Tokenization: Using Pre-trained Tokenizers (e.g., google/electra-small-discriminator).
3. Data Split: 80% Training / 20% Testing.

📊 Model Evaluation & Results

We experimented with 4 models using GPU T4 acceleration. ELECTRA was selected as the best model due to its superior training speed with comparable accuracy.

Model	Accuracy	Training Time	Status
ELECTRA	89%	2 min 59 sec	✅ Selected
DistilBERT	90%	7 min 4 sec	-
RoBERTa	89%	14 min 54 sec	-
DeBERTa	90%	18 min 17 sec	-

ELECTRA Performance Report

              precision    recall  f1-score   support

    negative       0.75      0.70      0.72       115
     neutral       0.91      0.87      0.89       158
    positive       0.91      0.94      0.93       427

    accuracy                           0.89       700
   macro avg       0.86      0.84      0.85       700
weighted avg       0.88      0.89      0.88       700