Aspect-Based Sentiment Analysis for Game Comments

This is a BERT-based classifier for performing aspect-based sentiment analysis (ABSA) on user comments about video games.
Each prediction considers both the aspect (topic/feature being discussed) and the comment text as inputs, and classifies the sentiment into 3 categories:

Negative
Neutral
Positive

🔍 How the Model Works

The model input consists of two segments:

Aspect (the topic whose sentiment you want to evaluate)
Comment Text (the actual user comment)

These are separated by a [SEP] token according to the BERT input format.
The model is trained using Focal Loss to handle class imbalance.

📊 Dataset

The dataset used for training consists of user comments on video games with the following columns:

comment_text
aspect
sentiment (0 = Negative, 1 = Neutral, 2 = Positive)
Dataset link : https://huggingface.co/datasets/alwanrahmana/Aspect-based-sentiment-analysis

📈 Performance

The model was trained using 5-Fold Cross Validation.
Evaluation metrics include accuracy and F1-score, with per-aspect breakdowns.

🚀 How to Use

Install dependencies:

pip install transformers torch