StyleNet — Neural Text Style Transfer

A neural sequence-to-sequence model for controllable text style transfer, built with GRU-based encoder-decoder architecture. Given an input sentence, StyleNet rewrites it in a target style (e.g., formal to informal, factual to stylistic) while preserving the core meaning.

What is Text Style Transfer?

Text style transfer is the task of changing the style of a piece of text — its tone, formality, sentiment, or writing manner — while keeping the content the same. This is a challenging NLP problem that sits at the intersection of:

• Natural Language Generation — producing fluent, coherent output
• Controllable Generation — steering the output towards a target style
• Sequence Modeling — understanding and transforming language at the sentence level

Architecture

Input Sentence
      |
  [Encoder]  — GRU-based bidirectional encoder
      |
  [Content Representation]
      +
  [Style Vector]  — learned style embedding
      |
  [Decoder]  — GRU-based autoregressive decoder
      |
 Output Sentence (in target style)

• Encoder: Bidirectional GRU that reads the input sentence and produces a fixed-size content representation.
• Style Embedding: A learned vector for each style class, concatenated with the content representation.
• Decoder: GRU decoder that generates the output token by token, conditioned on content + style.
• Training: Trained on parallel and non-parallel corpora using reconstruction loss and style classifier feedback.

Features

• GRU-based sequence-to-sequence architecture
• Controllable style conditioning via learned embeddings
• Supports Bengali and English corpora
• Configurable style classes (formal/informal, factual/stylistic)
• Attention mechanism for better content preservation

Requirements

pip install torch numpy pandas tqdm

Usage

# Train the model
python train.py --data_path data/ --style formal --epochs 50

# Generate styled text
python generate.py --input "The weather is bad today." --target_style formal
# Output: "Today's meteorological conditions are unfavorable."

Results

The model is evaluated on:

• BLEU Score — measures content preservation
• Style Classifier Accuracy — measures how well the target style is achieved
• Perplexity — measures fluency of generated text

Related Work

This implementation is inspired by:

• Style Transfer from Non-Parallel Text by Cross-Alignment (Shen et al., 2017)
• Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer (Li et al., 2018)
• The original StyleNet paper for factual text generation

Author

Nishat Tasnim Mahi — AI/ML Researcher | NLP | Computer Vision | Multimodal Learning

GitHub