MIKA, a Chatbot with Sentimental Responses Generated by Cyclegan and Reinforcement Learning

The tools of social interaction, like chatbots, are question-answer systems that enable conversations between a non-human and human entity, but in certain areas, this tools need to understand questions and generate answers with components of sentiment or emotion. The services of emotional care, such as, tutoring in the higher education centers or universities could use technological tools to fulfill the human functions of a psychologist, therefore, it is necessary that the intelligent model underlying this tools, like a chatbot, it should be generate sentimental responses like a human would response in this kind of areas, where the sentiment is a relevant factor. The purpose of this research was to develop a chatbot based on a model with artificial intelligence techniques that generates sentimental responses proposed by others authors using a corpus in English translated to Spanish. Thus, a deep learning architecture based on cycleGAN and reinforcement learning was modeled and trained, using the dialogue corpus Twitter Chat Log that have had 3.7 thousand pairs of validated dialogues that have been translated into Spanish during the training and testing stages. The model reached an accuracy of more than 60% by generating responses with sentiment adequately with possible improvements to reach 70%. Finally, the chatbot that include the trained model as a response generator engine, has been evaluated by some humans, where it reached 65 % acceptance and shows that the sentimental component of some responses still needs to be improved because they did not present semantic coherence for a human.