Evergreen Ballroom Prefuction
Markovian Magic: A Suggestion Engine with Nothing to Hide (306738)Ruchi Asthana, IBM
*Jennifer Alexis Mallette, IBM
Keywords: Natual Language Processing, Machine Leanring, IBM Watson, Predictive Analytics, Suggestion Engine, Artificial Intelligence, Makov Model, Chatbot
Chatbots are an excellent way to engage with users on a webpage. At IBM the conversational guidance provided by live agents is an extremely important part of the customer journey. Since chatbots are rule based machines trained on intents and entities to provide responses, we built a question suggestion engine that could help chatbots guide conversations the way a human representative might. We used intent classification to identify the current intention of the user and the direction of user conversations. Watson Assistant provided the framework for intent identification, and a Markov Model provided the framework for intent and buyer stage prediction. Most suggestion engines rely on Neural Networks or Deep Learning models which leverage Hidden Layers. Our approach uses a simple Markov Model allowing the training process to be more transparent. Due to this transparency, users can interact more with the data and learn direct conversational trends easily. Using this framework, 36% of test passed on a random test sample of 25% of live-chat data. This means for every 1 in 2.7 questions asked we were able to predict and suggest a question the user would later ask in the conversation.