This repository provides a TensorFlow implementation of a seq2seq chatbot, aiming to reproduce the results of Google's "A Neural Conversational Model." It's suitable for researchers and developers interested in building deep learning-based conversational agents.

How It Works

The project utilizes a recurrent neural network (RNN) with an encoder-decoder architecture, specifically LSTMs, for sentence prediction. It supports multiple dialogue corpora, including Cornell Movie Dialogs, OpenSubtitles, and custom data, with options for pre-trained word embeddings to accelerate training.

Quick Start & Requirements

• Install: pip3 install -r requirements.txt
• Prerequisites: Python 3.5, TensorFlow (v1.0 tested), NumPy, CUDA (for GPU), NLTK, tqdm. Additional dependencies for web interface: Django, Channels, Redis.
• NLTK Data: python3 -m nltk.downloader punkt
• Web Interface Setup: Requires CHATBOT_SECRET_KEY export, Django migrations, and Redis server.
• More Info: Installation, Docker, Web Interface

Highlighted Details

• Supports multiple dialogue corpora (Cornell, OpenSubtitles, Ubuntu, etc.) and custom data.
• Includes a web interface for interactive chatting.
• Offers vocabulary filtering and size limiting for performance optimization.
• Allows visualization of computational graphs with TensorBoard.

Maintenance & Community

The project appears to be a personal implementation with limited recent activity. Links to pre-trained models are noted as not working.

Licensing & Compatibility

The README does not explicitly state a license.

Limitations & Caveats

The project is based on TensorFlow v1.0 and may require adjustments for newer versions. Pre-trained model links are broken. Training can be slow, and initial results with shorter sequences are basic, with longer sequences requiring significant training time and still producing evasive answers or exhibiting gender/tense confusion. Overfitting is a noted issue.