AI Chatbot from Scratch - Intent-Based NLP Project

Academic AI Project | Computer Engineering | CHARUSAT
Course: Natural Language Processing / AI (2022)
Student: Sahil Patel (20CE101)
Approach: Intent Matching + Deep Learning
Model: Neural Network (TensorFlow/Keras)

---

📚 Project Overview

An intent-based conversational AI chatbot built from scratch using Natural Language Processing (NLP) and Deep Learning. The bot understands user queries through pattern matching and responds intelligently using a trained neural network.

Key Features:

• ✅ Intent Classification - Understands user intent from messages
• ✅ NLP Processing - Tokenization, lemmatization, bag-of-words
• ✅ Deep Learning - Custom neural network (128-64 neurons)
• ✅ JSON-based Training - Easy-to-extend intent patterns
• ✅ Real-time Responses - Interactive command-line chat
• ✅ Pre-trained Model - Ready to use chatbotmodel.h5

---

🎯 How It Works

Architecture Pipeline:

User Input → Tokenization → Lemmatization → Bag of Words → Neural Network → Intent Prediction → Response Selection

1. Training Phase (training.py)

• Load intents from JSON
• Tokenize patterns using NLTK
• Lemmatize words (WordNetLemmatizer)
• Create bag-of-words representation
• Build neural network (Sequential model)
• Train on patterns-intent pairs
• Save model as chatbotmodel.h5

2. Inference Phase (chatbot.py)

• Load trained model
• Accept user input
• Process through NLP pipeline
• Predict intent with confidence threshold
• Select random response from matching intent
• Display to user

---

📂 Repository Structure

Chatbot-from-Scratch/
├── chatbot.py              # Main chatbot runtime (83 lines)
├── training.py             # Model training script (81 lines)
├── intents.json            # Intent patterns and responses (120 lines)
├── chatbotmodel.h5         # Trained model (174 KB)
├── words.pkl               # Vocabulary pickle file
├── classes.pkl             # Intent classes pickle file
├── requirements.txt        # Python dependencies
├── .gitignore              # Git ignore patterns
├── LICENSE                 # MIT License
└── README.md               # This file

---

🚀 Getting Started

Prerequisites

Python 3.7+
TensorFlow 2.x
NLTK

Installation

# Clone repository
git clone https://github.com/patelsahil2k03/Chatbot-from-Scratch.git
cd Chatbot-from-Scratch

# Install dependencies
pip install -r requirements.txt

# Download NLTK data
python -c "import nltk; nltk.download('punkt'); nltk.download('wordnet'); nltk.download('omw-1.4')"

Quick Start

Option 1: Use Pre-trained Model

# Run chatbot with existing model
python chatbot.py

Option 2: Train from Scratch

# Train new model
python training.py

# Run chatbot
python chatbot.py

---

💬 Usage Example

$ python chatbot.py
GO! BOT IS RUNNING !
> Hello
Hi there, how can I help?

> What do you sell?
We sell coffee and tea

> Tell me a joke
Why did the hipster burn his mouth? He drank the coffee before it was cool.

> Thanks
Happy to help!

> Bye
See you later, thanks for visiting

---

🧠 Technical Details

Neural Network Architecture

model = Sequential()
model.add(Dense(128, input_shape=(len(train_x[0]),), activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(64, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(len(train_y[0]), activation='softmax'))

Layers:

• Input Layer: Variable size (vocabulary length)
• Hidden Layer 1: 128 neurons + ReLU + 50% Dropout
• Hidden Layer 2: 64 neurons + ReLU + 50% Dropout
• Output Layer: Softmax (number of intent classes)

Training Configuration:

• Optimizer: SGD (Stochastic Gradient Descent)
• Learning Rate: 0.01
• Momentum: 0.9 (Nesterov)
• Loss: Categorical Cross-entropy
• Epochs: 200
• Batch Size: 5

---

📝 Intent Configuration (intents.json)

Available Intents (10 total):

1. greeting - Hi, Hello, Hey
2. goodbye - Bye, See you later
3. thanks - Thank you, Thanks a lot
4. items - What do you sell?
5. payments - Payment methods
6. delivery - Shipping information
7. funny - Tell me a joke
8. time period - Internship duration (custom)
9. hours - Business hours
10. name - Bot name (Flash)

Adding New Intents:

{
  "tag": "your_intent",
  "patterns": [
    "User question 1",
    "User question 2",
    "User question 3"
  ],
  "responses": [
    "Bot response 1",
    "Bot response 2"
  ]
}

After adding, retrain the model:

python training.py

---

🛠️ Technologies Used

Core Libraries:

• TensorFlow/Keras - Deep learning framework
• NLTK - Natural language processing
• NumPy - Numerical computations
• Pickle - Model serialization
• JSON - Intent data storage

NLP Techniques:

• Tokenization - Split text into words
• Lemmatization - Reduce words to base form
• Bag of Words - Convert text to numerical vectors
• Intent Classification - Categorize user messages

Model Components:

• Sequential Neural Network - Feedforward architecture
• Dense Layers - Fully connected neurons
• Dropout - Regularization to prevent overfitting
• Softmax Activation - Multi-class probability distribution

---

📊 Code Breakdown

training.py (Model Training)

Key Functions:

1. Load intents from JSON
2. Tokenize and lemmatize patterns
3. Create vocabulary (words) and classes (intents)
4. Generate bag-of-words training data
5. Build and compile neural network
6. Train model (200 epochs)
7. Save model and pickles

Output Files:

• chatbotmodel.h5 - Trained model
• words.pkl - Vocabulary list
• classes.pkl - Intent classes

---

chatbot.py (Inference)

Key Functions:

clean_up_sentence(sentence)
# Tokenizes and lemmatizes input

bag_of_words(sentence)
# Converts sentence to numerical vector

predict_class(sentence)
# Predicts intent with confidence threshold (0.25)

get_response(intents_list, intents_json)
# Selects random response from matched intent

Workflow:

1. Load model and pickles
2. Enter infinite chat loop
3. Process user input through NLP pipeline
4. Predict intent (with 25% confidence threshold)
5. Select and display response

---

🎓 Academic Context

Course: Natural Language Processing / Artificial Intelligence
Semester: 4th/5th Semester (2022)
Institution: CHARUSAT - CSPIT
Program: B.Tech Computer Engineering
Student ID: 20CE101

Purpose: Build conversational AI from scratch to understand NLP and neural networks.

---

🚀 From Basics to Production (2022 → 2026)

This repository represents foundational NLP/AI learning from 2022.

Journey Since Then:

• 🤖 Built production chatbots with LangChain and advanced LLMs
• 📊 Achieved 98%+ accuracy in AI models (production systems)
• 🔬 Published 2 SCOPUS-indexed papers on AI/ML
• 🏆 Top 10 Finalist in AI-Manthan Hackathon
• ☁️ Deployed 50+ AWS Lambda AI functions
• 💼 Associate Software Engineer building AI solutions at Digiflux

Current Expertise: LangChain, GPT, BERT, Transformers, Production NLP
Portfolio: patelsahil2k03.github.io

---

💡 Key Learnings

NLP Concepts:

• ✅ Tokenization - Text preprocessing fundamentals
• ✅ Lemmatization - Word normalization techniques
• ✅ Bag of Words - Text vectorization
• ✅ Intent Recognition - Classification problems

Deep Learning:

• ✅ Neural Networks - Architecture design
• ✅ Backpropagation - Training process
• ✅ Dropout - Regularization techniques
• ✅ Softmax - Multi-class classification

Software Engineering:

• ✅ Modular Design - Separate training and inference
• ✅ Model Persistence - Saving/loading models
• ✅ JSON Configuration - Data-driven approach
• ✅ Error Handling - Confidence thresholds

---

🔧 Customization Guide

1. Change Bot Personality

Edit intents.json responses to match desired tone.

2. Add Domain-Specific Intents

Add new intent blocks for your use case (e-commerce, support, etc.)

3. Improve Accuracy

• Increase epochs (200 → 500)
• Add more training patterns per intent
• Tune hyperparameters (learning rate, neurons)

4. Multi-language Support

• Use language-specific lemmatizers
• Translate intents.json

5. GUI Integration

Wrap chatbot.py logic in Flask/Streamlit for web interface

---

📈 Performance Metrics

• Training Time: ~2-5 minutes (200 epochs)
• Model Size: 174 KB (lightweight)
• Response Time: < 100ms per query
• Confidence Threshold: 25% (adjustable)
• Intent Accuracy: Depends on training data quality

---

🤝 Contributing

This is an academic learning project. Feedback welcome!

Enhancement Ideas:

• Add more intents (100+ for robust bot)
• Implement context handling (conversation memory)
• Add sentiment analysis
• Integrate with web/mobile interface
• Add voice input/output

---

📜 License

MIT License - See LICENSE file for details.

---

📧 Contact

Sahil Patel
Email: patelsahil2k03@gmail.com
Portfolio: patelsahil2k03.github.io
GitHub: @patelsahil2k03
LinkedIn: sahil-patel-581226205

---

🌟 Acknowledgments

• CHARUSAT CSPIT - NLP/AI curriculum
• TensorFlow Team - Excellent deep learning library
• NLTK - Comprehensive NLP toolkit
• AI Community - Inspiration and resources

---

📚 References

• NLTK Documentation: https://www.nltk.org/
• TensorFlow Keras: https://www.tensorflow.org/guide/keras
• Intent-based Chatbot tutorials
• Bag-of-Words model explanations

---

From basic intent matching to building production LLM applications - this project marks the beginning of my AI journey.
Last Updated: March 2026