Emotion Classifier — COMP 4107 Final Project

Attention-Based Emotion Classification with Template-Based Decision Support

Ziyi Jiang - 101266200 Zhimei Li - 101258414

Overview

This project builds an emotion classification system for short English texts. Given a sentence, the system predicts one of six emotions, highlights the words that influenced the prediction using attention weights, and provides a short piece of structured judge-style guidance based on the result.

Two models are compared:

BiGRU + Masked Pooling (baseline) BiGRU + Attention (proposed)

The attention model achieves similar classification performance to the baseline but adds interpretability — you can see which words the model focused on when making a prediction. 6 emotion classes: sadness, joy, love, anger, fear, surprise

Dataset

We use the Kaggle Emotions dataset.

About 400k short English texts Each text has one label (6 classes)

The data is slightly imbalanced. For example, joy and sadness appear more often than love and surprise.

Preprocessing

Before training, we do a few simple steps:

1. convert text to lowercase
2. remove URLs and mentions
3. clean special characters
4. tokenize text into words
5. pad or truncate to fixed length

Models

1. BiGRU (Baseline)

• word embeddings
• bidirectional GRU
• mean + max pooling
• linear classifier

2. BiGRU + Attention

• same BiGRU encoder
• additive attention layer
• weighted sum of hidden states
• classifier

The attention layer helps the model focus on important words.

How to Run

• Setup: pip install -r requirements.txt

• Train both models: python src/train.py --model baseline --epochs 30 --data data/emotions.csv python src/train.py --model attention --epochs 30 --data data/emotions.csv

• Evaluate and compare: python src/evaluate.py --compare --data data/emotions.csv

• Run the interactive demo: python src/demo.py --model attention

• Then type a sentence like: I feel really sad and alone I am so frustrated with everything right now I am scared of what will happen next

Project Structure

src/ preprocess.py — text cleaning, vocabulary, train/val/test split, guidance templates dataset.py — PyTorch Dataset and DataLoader model.py — BiGRUBaseline and BiGRUAttention model definitions train.py — training loop with early stopping evaluate.py — test set evaluation, confusion matrix, model comparison predict.py — single-text prediction logic demo.py — interactive command-line demo utils.py — shared utilities (metrics, plotting, checkpointing)

data/ emotions.csv — Kaggle emotions dataset (not tracked in git)

outputs/ — model checkpoints, plots (not tracked in git)