Investigating Deep Learning for Predicting Multi-linguistic Interactions with a Chatterbot

Deep Learning (DL) becomes a mainstream technique for Artificial Intelligence (AI) machine learning because of its success in performing many tasks, such as image recognition, speech interpretation, language prediction and translation. We are investigating the underlying principles of DL Neural Networks (NN) to design optimal DL NN for predicting human multi-linguistic conversations with a chatterbot. This research attempts to tackle the well-known open problem of finding optimal NN designs for data of various characteristics. We are in particular focusing on Recurrent Neural Networks (RNN) models with time progression, which takes into consideration the results from the previous steps plus the current input to predict the next step, i.e. it ‘remembers’ what it has previously learnt. Through the experiments of tuning an RNN to achieve an optimal performance in terms of accuracy and training time, we found that characteristics such as word counts and layers of neurons could affect the training performance. We applied the tuned optimal model to a game implementation, inspired by IBM Watson, where users can guess the words to be generated by a computer, called “Beat AI” to have human predict the machine prediction.