Deep Learning for Session Aware Conversational Agents

In the last 2 years the state of NLP research has made a huge step forward. Since the release of ELMo, a new race for the leading scoreboards of all the main linguistic tasks has begun. Several models came out every 2 months achieving promising results in all the major NLP applications, from QA to text classification, passing through NER. These great research discoveries coincide with an increasing trend for voice and textual technologies in the customer care market. One of the next biggest challenge will be the handling of multi-turn conversations, a types of conversation which differs from single-turn by the presence of the concept of session. A session is a set of related QA between the user and the agent to fulfill a single user request. A conversational agent has to be aware about the session to effectively carry on the conversation and understand when the goal can be achieved. The proposed work focused on three main parts: i) the study of the state of the art deep learning techniques for NLP ii) the presentation of a model, MTSI BERT (Multi Turn Single Intent BERT), using one of such NLP milestones in a multi-turn conversation scenario iii) The study of a real case scenario. The work takes in consideration both Recurrent Neural Networks and attention based models, as well as word embedding such as Word2Vec and Glove. The proposed model, based on BERT and biLSTM, achieves promising results in conversation intent classification, knowledge base action prediction and detect of end of dialogue session, to determine the right moment to fulfill the user request. The study about the realization of PuffBot, an intelligent chatbot to support and monitor asthma suffering children, shows how this type of technique could be an important piece in the development of future chatbot.