A Behavioral Chatbot Using Encoder-Decoder Architecture : Humanizing conversations

Though there are so many ways of building conversational chatbots, they lack the human touch and sound very robotic. We don’t have any chatbots that aim to imitate even a speck of a personality or human-like traits, while we very well possess everything that we need in terms of data and computation. This project aims to make both efficient and human-like chatbot using the modern encoder-decoder architecture. There are many frameworks and libraries available to develop AI-based chatbots including program-based, rule-based and interface-based. But they lack the flexibility in developing real dialogues and understanding humans. The popular chatbot models don’t aim to hold conversations that imitate real human-like interactions. Current chatbots employ a rule-based approach, basic machine learning algorithms, or a retrieval-based strategy that does not provide humanized outputs, i.e., these chatbots are incapable of producing engaging dialogues. In this paper, we tried to develop a Behavioural chatbot, using modern deep learning techniques like Seq2Seq aiming to develop chatbots to understand humans and make some situation agnostic conversations that remind us of our desirable personalities. This chatbot model was trained on real human conversations extracted out of Hollywood movies, hence the dataset possesses around 2.3 lakh truly organic dialogues. The model was trained with ~4.2 million parameters, 250 epochs and attained an accuracy of 95%. This chatbot is capable of displaying subtle sarcasm and also tries to be funny at times, thanks to the dramatic Hollywood dialogue writers.