nlp_for_journal

BERT + LSTM Sentiment Analysis Model

This repository contains a custom fine-tuned BERT + LSTM model for sentiment classification, trained on journal-style emotional writing. The model is designed to detect nuanced emotional states and works well for reflective, personal text (e.g., journaling, social media, diary entries).

---

🧠 Model Architecture

• Base Encoder: bert-base-uncased (Hugging Face Transformers)
• Sequence Modeling: One-layer LSTM to capture sentence-level emotional flow
• Classification Layer: Fully connected layer with softmax
• Loss Function: CrossEntropyLoss
• Optimizer: AdamW

---

📦 Files Included

---

🧾 Classes

The model was trained to classify the following emotional labels:

• Positive
• Negative
• Neutral

---

🚀 How to Use

#pip install torch transformers scikit-learn #if not already installed, remove the first comment

import torch
import pickle
from transformers import BertTokenizer

# Load model
model = torch.load("bert_lstm_sentiment_model.pt", map_location=torch.device('cpu'))
model.eval()

# Load tokenizer and label encoder
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
with open("label_encoder.pkl", "rb") as f:
    label_encoder = pickle.load(f)

# Sample input
text = "I feel empty and overwhelmed today."

# Tokenize
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
with torch.no_grad():
    outputs = model(**inputs)
    predicted_class = torch.argmax(outputs.logits, dim=1).item()
    emotion = label_encoder.inverse_transform([predicted_class])[0]

print(f"Predicted Emotion: {emotion}")