Twitter Emotion Classification using BERT

Classify short tweets into six emotions using a fine-tuned BERT model:

• sadness
• joy
• love
• anger
• fear
• surprise

Features

• BERT fine-tuning with Trainer
• Stratified train/validation split
• Weighted & macro F1 (plus accuracy)
• Early stopping (best model restored)
• Dynamic padding via DataCollatorWithPadding
• Reproducible runs (set_seed)

Dataset

Uses https://huggingface.co/datasets/dair-ai/emotion
Local CSV also supported with columns: text, label in {sadness, joy, love, anger, fear, surprise}.

Requirements

pip install torch transformers datasets pandas numpy scikit-learn matplotlib


git clone https://github.com/your-username/bert-twitter-emotion.git
cd bert-twitter-emotion
jupyter notebook notebooks/BERT_Twitter_Emotion_Classification.ipynb

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

label_dict = {'sadness':0,'joy':1,'love':2,'anger':3,'fear':4,'surprise':5}
id2label = {v:k for k,v in label_dict.items()}

mdl = AutoModelForSequenceClassification.from_pretrained("models/bert-emotion")
tok = AutoTokenizer.from_pretrained("models/bert-emotion")
def predict(texts):
    enc = tok(texts, padding=True, truncation=True, return_tensors="pt")
    with torch.no_grad():
        probs = mdl(**enc).logits.softmax(-1)
        ids = probs.argmax(-1).tolist()
    return [(t, id2label[i], float(probs[j, i])) for j,(t,i) in enumerate(zip(texts, ids))]