CNN-BiLSTM Classifier

A hybrid CNN+BiLSTM architecture for product classification with attention mechanism, implemented in PyTorch.

Features

• Multi-kernel CNN: Captures local n-gram features with different kernel sizes (2, 3, 4, 5)
• Bidirectional LSTM: Captures sequential dependencies with soft attention mechanism
• Ad-hoc Features: Extracts statistical and linguistic features from text
• Word2Vec Embeddings: Pre-trained word embeddings for better text representation
• Easy Integration: Simple API for training and inference
• HuggingFace Hub Support: Load and save models to/from HuggingFace Hub

Installation

Install from PyPI:

pip install cnn-bilstm-classifier

Or install from source:

git clone https://github.com/turgutguvercin/cnn-bilstm-classifier.git
cd cnn-bilstm-classifier
pip install -e .

Quick Start

Loading a Pre-trained Model

from product-classifier import CNNBiLSTMInference

# Load from HuggingFace Hub
model = CNNBiLSTMInference.from_pretrained("turgutguvercin/product-classifier-v1")

# Make predictions
predictions = model.predict([
    "Yataş Bedding BAMBU Yorgan (%20 Bambu) 300 Gr.",
    "Arji Ev ve Ofis Çalışma Sandalyesi Bilgisayar Koltuğu"
], top_k=3)

for i, pred in enumerate(predictions):
    print(f"Text {i+1} predictions:")
    for label, score in pred:
        print(f"  → {label}: {score:.4f}")

Training a New Model

from product_classifier import CNNBiLSTMTrainer
import pandas as pd

# Load your data
df = pd.read_csv("your_data.csv")

# Initialize trainer
trainer = CNNBiLSTMTrainer()

# Train the model
results = trainer.train(
    df,
    text_column="product_name",
    label_column="category",
    config={
        'batch_size': 128,
        'epochs': 20,
        'embedding_dim': 512,
        'cnn_filters': 128,
        'lstm_hidden': 256
    }
)

# Save the model
trainer.save_model("./my_model")

Command Line Interface

Train a model:

cnn-bilstm-train --data data.csv --text-column name --label-column category --output-dir ./model

Make predictions:

cnn-bilstm-predict --model-path ./model --text "Product name to classify"

Model Architecture

The CNN+BiLSTM classifier combines three main components:

1. Multi-kernel CNN Branch:

   • Convolutional layers with kernel sizes [2, 3, 4, 5]
   • Max pooling over sequence dimension
   • Captures local n-gram patterns

2. BiLSTM Branch:

   • Bidirectional LSTM layers
   • Soft attention mechanism
   • Captures long-range dependencies

3. Ad-hoc Features Branch:

   • Statistical features (length, character ratios)
   • Linguistic features (word count, symbol count)
   • Word length histograms

All branches are concatenated and fed to a final classification layer.

API Reference

CNNBiLSTMInference

Main class for model inference.

Methods

• from_pretrained(repo_id): Load model from HuggingFace Hub
• from_local(model_dir): Load model from local directory
• predict(texts, top_k=1): Predict categories for input texts
• get_model_info(): Get model architecture information

Example

# Load model
model = CNNBiLSTMInference.from_pretrained("username/model-name")

# Single prediction
prediction = model.predict("Product title", top_k=3)

# Batch prediction
predictions = model.predict(["Title 1", "Title 2"], top_k=1)

# Model info
info = model.get_model_info()
print(f"Vocabulary size: {info['vocab_size']}")

CNNBiLSTMTrainer

Class for training new models.

Methods

• train(df, text_column, label_column, config): Train a new model
• save_model(output_dir): Save trained model
• save_to_hub(repo_name): Upload model to HuggingFace Hub

AdHocFeatureExtractor

Extracts statistical and linguistic features from text.

Features Extracted

1. Title length
2. Uppercase character rate
3. Alphabetic character rate
4. Digit character rate
5. Space count and rate
6. Word count
7. Maximum word length
8. Unique word rate
9. Symbol count
10. Word length histogram (5 bins)

Configuration

Default training configuration:

config = {
    'batch_size': 128,
    'epochs': 20,
    'learning_rate': 0.001,
    'embedding_dim': 512,
    'cnn_filters': 128,
    'lstm_hidden': 256,
    'max_length': 50
}

Requirements

• Python >= 3.8
• PyTorch >= 1.9.0
• NumPy >= 1.21.0
• SafeTensors >= 0.3.0
• HuggingFace Hub >= 0.15.0
• Gensim >= 4.0.0
• scikit-learn >= 1.0.0
• tqdm >= 4.60.0

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

1. Fork the repository
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If you use this work in your research, please cite:

@software{product_classifier,
    title={CNN-BiLSTM Classifier: Hybrid Architecture for Product Classification},
    author={Turgut Guvercin},
    year={2025},
    url={https://github.com/turgutguvercin/cnn-bilstm-classifier},
    note={Implementation based on the methodology from Suzuki et al. (2018)}
}

@inproceedings{suzuki2018cnn,
    title={Convolutional Neural Network and Bidirectional LSTM Based Taxonomy Classification Using External Dataset at SIGIR eCom Data Challenge},
    author={Suzuki, Shogo D. and Iseki, Yohei and Shiino, Hiroaki and Zhang, Hongwei and Iwamoto, Aya and Takahashi, Fumihiko},
    booktitle={Proceedings of ACM SIGIR Workshop on eCommerce (SIGIR 2018 eCom Data Challenge)},
    year={2018},
    month={July},
    address={Ann Arbor, Michigan, USA},
    publisher={ACM},
    pages={1--5},
    url={https://sigir-ecom.github.io/ecom2018/ecom18DCPapers/ecom18DC_paper_1.pdf},
    note={Original paper describing CNN+BiLSTM architecture for product taxonomy classification}
}

Acknowledgments

• Based on "Convolutional Neural Network and Bidirectional LSTM Based Taxonomy Classification"
• Built with PyTorch and HuggingFace ecosystem
• Special thanks to the open-source community