Railway AI Assistant – Conversational State Tracking

Overview

This project implemented a conversational railway assistant using a fine-tuned T5-Base Transformer model. The system converted multi-turn railway conversations into structured JSON states by extracting user intent, destination, preferences, constraints, and conversation status.

Features

• Intent classification (search_train, platform_info, check_disruptions)
• Multi-turn context retention
• User correction handling
• Constraint and preference extraction
• Structured JSON state generation
• Lightweight frontend for inference

Dataset

• Synthetic railway conversation dataset
• ~8,500 conversation samples
• Conversation-to-JSON state mapping

Model

• T5-Base (Encoder-Decoder Transformer)
• Fine-tuned using Hugging Face Transformers
• Trained on Google Colab GPU with FP16 mixed precision

Evaluation

The model was evaluated using:

• Exact JSON Match Accuracy
• Intent Accuracy
• Destination Accuracy
• Constraint Accuracy
• Status Accuracy
• ROUGE-1
• ROUGE-2
• ROUGE-L

Example

Input

User: Cheapest train to Manchester tomorrow

Output

{
  "intent": "search_train",
  "destination": "Manchester",
  "time_context": "tomorrow",
  "preference": "cheapest",
  "status": "new_search"
}

Technologies Used

• Python
• PyTorch
• Hugging Face Transformers
• T5-Base
• Pandas
• NumPy
• Google Colab

Project Structure

├── notebook.ipynb
├── railway_dataset.csv
├── saved_model/
├── frontend/
├── README.md
└── requirements.txt

Outcome

The solution successfully demonstrated conversational state tracking by maintaining context, processing user updates, and generating structured railway states suitable for downstream journey-planning systems.