Authors: Varshith Gude
Framework: PyTorch
Date: 2025

This project presents a sequence-to-sequence (Seq2Seq) model capable of converting numeric strings into their corresponding word representations. The system efficiently maps sequences of digits (0–9) into sequences of English words (e.g., 123 → one two three) using a GRU-based encoder-decoder architecture. This approach demonstrates the feasibility of applying neural sequence models to structured symbolic translation tasks.

Automatic conversion of numeric data into natural language is essential in applications such as:

• Voice assistants reading numbers aloud.
• Financial document processing.
• Educational tools for learning numerical literacy.

The task highlights sequence modeling challenges with variable-length inputs and outputs, providing a controlled environment to study encoder-decoder architectures with greedy decoding.

• Encoder: GRU-based network that encodes the input numeric sequence into a hidden representation.
• Decoder: GRU-based network that generates the word sequence conditioned on the encoder hidden states.
• Seq2Seq Framework: Connects encoder and decoder for end-to-end sequence translation.

Digit-to-word mapping:

digit_to_word = {
'0': 'zero', '1': 'one', '2': 'two', '3': 'three', '4': 'four', '5': 'five', '6': 'six', '7': 'seven', '8': 'eight', '9': 'nine'
}

--

• config.py – Project configuration and hyperparameters.
• dataset.py – Dataset generation, save/load utilities, and PyTorch Dataset class.
• generate_dataset.py – Script to generate training data.
• train.py / main.py – Training script and/or interactive number converter.
• models.py – Encoder, decoder, and Seq2Seq model definitions.
• train_utils.py – Training helper functions.
• vocab.py – Input/output vocabularies and token mappings.
• inference.py – Greedy decoding function for inference.

• Loss function: Cross-entropy between predicted and true word sequences.
• Optimization: Adam optimizer, learning rate 0.001.
• Batch size: 128
• Embedding size: 64
• Hidden size: 128
• Teacher forcing ratio: 0.5
• Epochs: 20

• Greedy decoding: Generates exactly as many words as the input digits, avoiding early stopping and <eos> token issues.

# Test The Model
python test.py