NLP Intent Parser — End-to-End Industrial Command Understanding

This project implements a complete intent parsing system for technician-style instructions in industrial and microgrid environments. It takes raw natural language like:

“check the inverter temperature and update the power limit to 20%”

…and outputs structured actions:

{
  "intent": "update_parameter",
  "target": "inverter",
  "parameter": {
    "name": "power_limit",
    "value": "20%"
  }
}

The system benchmarks three NLP modeling families:

1. TF-IDF + Linear SVM — baseline intent classification
2. LSTM / BiLSTM — multi-output classification (intent + target + parameter)
3. DistilBERT Token Classifier — end-to-end structured extraction

The goal: build a clean, production-style pipeline that demonstrates how traditional ML, classical deep learning, and modern transformers differ in capability and performance.

---

📌 Key Features

✓ Synthetic Dataset Generation

Because real technician logs are private and inconsistent, the project builds a controlled, balanced synthetic dataset covering:

• 10+ intents
• 15+ equipment targets
• 20+ parameter types
• Numeric, categorical, and percentage values

Flexible enough to extend or adapt to real operational logs later.

---

✓ Multi-Model Benchmarking

Each model solves the same set of tasks:

• Intent classification
• Target identification
• Parameter extraction

This allows for direct comparison between:

Model	Strengths	Weaknesses
TF-IDF + SVM	Fast, simple, solid baseline	No structured extraction
LSTM / BiLSTM	Multi-output, learns patterns	Needs hand-engineered preprocessing
DistilBERT	Best overall generalisation; robust extraction	Heavier, slower on CPU

---

✓ End-to-End Parsing Demo

A final pipeline demonstrates how raw text becomes structured output:

• Tokenisation
• Transformer inference
• Slot grouping
• Value extraction
• JSON-like final output

This is the most production-like part of the system and the highlight of the project.

---

✓ Clear, Modular Notebook Structure

The notebook is intentionally structured so that:

• Students can follow it
• Recruiters can evaluate it quickly
• Engineers can adapt it to production

Sections include:

1. Imports & Setup
2. Dataset Generation
3. Preprocessing
4. TF-IDF Baseline
5. LSTM / BiLSTM Models
6. DistilBERT Model
7. Unified Evaluation
8. End-to-End Demo
9. Model Comparison Summary

---

🏗 Architecture Overview

Raw Instruction
        ↓
   Preprocessing
        ↓
  Three Model Families
        ↓
Unified Evaluation Layer
        ↓
 End-to-End Parser Demo
        ↓
 Structured JSON Output

---

🧪 Example Usage

Run the final cell in the notebook:

parse_instruction("reset the inverter frequency to 50hz")

Output:

{
  "intent": "reset",
  "target": "inverter",
  "parameter": {
    "name": "frequency",
    "value": "50hz"
  }
}

---

📊 Model Performance Summary

While scores vary depending on the exact synthetic dataset, trends are consistent:

• TF-IDF + SVM performs well for intent only
• BiLSTM improves multi-output performance
• DistilBERT dominates structured extraction

Transformers are recommended if you want:

• robustness to grammar changes
• better understanding of technician lingo
• high accuracy on parameter extraction

---

⚙️ Installation

pip install -r requirements.txt

Or install notebook dependencies manually:

transformers
tensorflow
torch
pandas
numpy
scikit-learn
tqdm
matplotlib

---

📁 Repository Structure

.
├── e2e_intent_parser.ipynb
├── README.md
└── data/

---

🎯 Why This Project Matters

Industrial environments need interpretable AI — not black boxes.

Technicians speak in short, imperative commands with variable structure. This project shows how to build an AI system that can:

• parse real operator instructions
• extract actionable parameters
• integrate into predictive maintenance and control systems
• adapt across equipment types

It’s both a portfolio piece and a template for real deployments.

---

🚀 Next Steps

Planned improvements:

• Add CRF layer on top of DistilBERT
• Add error-recovery heuristics for incomplete queries
• Export a standalone Python library (intent_parser/)
• Production API (FastAPI + lightweight ONNX model)