udayallu / seq2seq-chatbot

Chatbot based Seq2Seq model with bidirectional rnn and attention mechanism with tensorflow, trained on Cornell Movie-Dialogs Corpus and deployed on a Flask Server

Seq2Seq-Chatbot

Introduction

This Chatbot is a TensorFlow implementation of Seq2Seq Mode. It make use of a seq2seq model RNN for sentence predictions. The chatbot is trained on Cornell Movie Dialogs Corpus on Conversation of lenghth between 5 to 2 words

Installation

To run this chabtot you will need:

• Python 3.5
• Tensorflow (1.13.1)
• Numpy
• Flask
• Matplotlib
• Dataset

Running

To chat with the trained model from Web Chat:

1. First install the required depedencies from Chat_UI_Web and run pip install -r requirment.txt
2. From the Chat_UI_Web directory run run.py python run.py
3. Open a browser to the URL indicated by the server console, This is typically: http://localhost:5000/

To chat with the trained model from Python Console:

1. First install the required depedencies from root directory and run pip install -r requirment.txt
2. From the root directory run python chat.py

List of Hyperparameters

• Batch_size to 128
• Rnn_size to 128
• Embed_size to 128
• Learning_rate to 0.0.1
• Epochs to 500
• Keep_prob to 0.75

For Training Model

To Train Model from Python Console:

1. Configure the config.py file for desired training hyperparameters
2. Run train.py python train.py
3. Make sure the datasets should be present in the datasets folder

Results

Model was able to train to Accuracy of 62% and Loss of 19% and after that it was overfitting

Accuracy

Loss

Chat Results

Human: who are you
Bot: girlfriend
Human: who am i
Bot: doctor livingston
Human: where are you from
Bot: southern california
Human: how old are you
Bot: nineteen or
Human: are you a boy
Bot: i am trying
Human: are you a girl
Bot: something like that
Human: do you drink
Bot: no thanks

Some Funny Result