Server-Less Rule-Based Chatbot Using Deep Neural Network

Sammanfattning: Customer support entails multi-faceted benefits for IT businesses. Presently, the business depends upon on conventional channels like e-mail, customer care and web interface to provide customer support services. However, with the advent of new developments in Scania IT, different IT business units is driving a shift towards automated chatbot solutions to provide flexible responses to the user's questions. This thesis presents a practical study of such chatbot solution for the company SCANIA CV AB, Södertälje. The objective of the research work presented in this thesis is to analyze several deep learning approaches in order to develop a chatbot prototype using serverless Amazon Web Services components. The proposed bot prototype includes two main Natural Language Understanding (NLU) tasks: Intent classification and Intent fulfilment. This is a two-step process, focusing first on Recurrent Neural Network (RNN) to perform a sentence classification (intent detection task). Then, a slot filling mechanism is used for intent fulfilment task for the extraction of parameters. The results from several neural network structures for user intent classification are analyzed and compared. It is found that the bidirectional Gated Recurrent units (GRU) were shown to be the most effective for the classification task. The concluded model is then deployed on the designed AWS stack. They demonstrate that the bot behaves as expected and it places more insistence on the structure of the neural network and word embeddings for future advancements in order to find an even better neural network structure.