How to train DeepQA on a custom dataset?

Follow the lightweight format instructions in data/lightweight, preprocess your conversations, and use the --corpus flag with main.py, but note that manual preprocessing and TensorFlow 1.0 compatibility may require adjustments.

DeepQA vs Seq2Seq in TensorFlow 2: which is better for learning?

DeepQA is good for understanding classic seq2seq chatbots with a focused codebase, but TensorFlow 2 offers updated tutorials and models; for modern practices, TF2 resources might be more relevant due to DeepQA's outdated framework.

Does DeepQA support attention mechanisms out of the box?

No, attention is listed as a potential improvement in the README, requiring manual code changes by replacing embedding_rnn_seq2seq with embedding_attention_seq2seq, which adds complexity for users.

What GPU do I need to train DeepQA effectively?

The README mentions training on a GeForce GT 740M with reduced parameters, but for better results, a higher-end GPU is recommended due to the computational intensity of seq2seq models and large corpora.

Can DeepQA handle multi-turn conversations?

By default, it processes questions independently, but the improvements section suggests feeding previous dialogue with tokens, though this requires retraining and may not capture long-term dependencies well.

Is the pretrained model for DeepQA still available to download?

No, the pretrained model links in the README are broken, and reliance on community contributions means users often need to train from scratch or find alternative sources.

Open-Awesome

Chatbot

Apache-2.0Python

A TensorFlow implementation of a neural conversational model (seq2seq) for building deep learning chatbots.

GitHub

2.9k stars1.2k forks0 contributors

What is Chatbot?

DeepQA is a TensorFlow-based implementation of a neural conversational model (seq2seq) that builds chatbots using deep learning. It replicates the approach from the Google research paper "A Neural Conversational Model," allowing training on dialogue corpora to generate human-like responses. The project provides tools for training, testing, and deploying chatbot models with support for multiple datasets and a web interface.

Target Audience

Machine learning practitioners, researchers, and developers interested in building and experimenting with neural network-based chatbots using seq2seq architectures and TensorFlow.

Value Proposition

It offers a reproducible, open-source codebase for neural conversational models with multi-corpus support, extensible design for improvements like attention, and a ready-to-use web interface for deployment.

Overview

My tensorflow implementation of "A neural conversational model", a Deep learning based chatbot

Use Cases

Best For

Experimenting with seq2seq models for dialogue generation
Training chatbots on custom conversation datasets
Learning neural conversational AI with TensorFlow
Deploying a self-hosted chatbot with a web interface
Researching improvements like attention or sampling in dialogue systems
Comparing performance across different dialogue corpora

Not Ideal For

Production applications needing scalable, real-time chatbots with high accuracy
Teams without dedicated GPU resources for training large neural networks
Developers seeking state-of-the-art transformer models like GPT for advanced dialogue

Pros & Cons

Pros

Multiple Corpus Support

Supports diverse datasets like Cornell Movie Dialogs and OpenSubtitles, enabling training on varied dialogue styles directly from the README instructions.

Extensible Research Codebase

Clear TensorFlow implementation allows easy experimentation with seq2seq improvements, such as adding attention or sampling mechanisms as highlighted in the improvements section.

Integrated Web Interface

Includes a Django-based UI with Redis for real-time deployment, providing a practical way to test chatbots interactively beyond command-line.

Pre-trained Embeddings Option

Allows use of external word embeddings to speed up training and enhance model performance, as documented in the data/embeddings section.

Cons

Outdated TensorFlow Dependency

Built on TensorFlow 1.0, which is deprecated and lacks support for modern features, making it incompatible with newer TensorFlow versions and optimizations.

Complex Setup and Maintenance

Requires manual installation of CUDA, Redis, and multiple Python dependencies, with broken pretrained model links adding to the setup hassle.

Limited Conversational Features

Lacks built-in attention or sampling mechanisms, leading to deterministic, often evasive responses and overfitting, as admitted in the results and improvements sections.

Frequently Asked Questions

Related Projects

Kubeflow

Machine Learning Toolkit for Kubernetes

Stars15,852

Forks2,695

Last commit17 days ago

Policy Gradient

Deep Learning and Reinforcement Learning Library for Scientists and Engineers

Stars7,379

Forks1,583

Last commit3 years ago

YOLO TensorFlow ++

Translate darknet to tensorflow. Load trained weights, retrain/fine-tune using tensorflow, export constant graph def to mobile devices

Stars6,139

Forks2,015

Last commit2 years ago

Sentence Classification with CNN

Convolutional Neural Network for Text Classification in Tensorflow

Stars5,686

Forks2,725