Chatbot-Reddit

Chatbot that is trained on Reddit dump of 2005. This is a word level model which predicts the next sentence based on question statement and context (comments and replies in this case). The model uses encoder/decoder architecture and teacher forcing.

Dependencies

• Keras (Tensorflow Backend)
• Numpy
• Scipy
• Jupyter

Usage

1. Download and install Jupyter Notebook and IPython kernel
2. run a Jupyter environment locally using jupyter notebook in the terminal
3. call load_model() to load pretrained model in /models dir or train model for atleast 30 epochs
4. change temperature of output(random sampling coefficient) in sample()
   
   (higher value of temperature = higher randomness, lower value of temperature preserves local structure but increases redundancy)
5. call make_inference() with question statements to generate outputs

Example Output

Question: life
Reply: joke not did apparently best non wrong usually when call love you seriously

References

• Keras Blog by Francois Chollet

TODO

• [] Train on dialog dataset for proper answer structure
• [] Clean dataset for stop words using nltk library
  
  

• The bot might say some offensive things(trained on reddit)