Studying Chatbot with Deep Learning

Abstract

In this research, we made an effort to improve current models to get better performance for automatically replying a human query. The project is to create an intelligent chatbot for Stratford University where students’ queries are required to be entertained by the system. In this research, we studied multiple methods of creating chatbot, e.g. rule-based and generative models. We have focused on using Generative models for Chatbot during our research. Further study helped us with devising a chatbot using Deep Learning Recurrent Neural Network (RNN) and seq2seq model. A seq2seq model consists of two back-to-back RNNs. One of them is Encoder and the other is decoder. Between these two RNNS, we have Long Short Term Memory (LSTM), which helps us to maintain the memory of conversation in the Chatbot. Our experiments helped to get the best parameters values for the best performance out of the model, both during training and testing. A number of parameters were tested and some of them were found very appealing with respect to the size of datasets. For evaluation of the model, we have used BLEU, a standard use to evaluate such models and have seen the effects of multiple datasets on this model. Future directions and ways of improvement are discussed in the end.