Dialog systems offer a natural means of communication with computers and other devices. Recently they have been experiencing rapid progress, mainly thanks to advances in machine learning.
Current dialog system research is split into two branches: (1) Task-oriented systems typically focus only on a few selected communication domains. Their ability to understand conversation is limited, which may lead to inappropriate responses. (2) Chit-chat systems do not have a specific goal other than a casual, but interesting and entertaining conversation with the user. However, they still lack controllability and ability to understand deeper aspects of language.
This project will focus on an underexplored field: combining the two types of systems. The goal will be to develop an efficient statistical model that will be able to accomplish tasks (e.g., book accommodation in a hotel), but also to engage in a casual conversation, i.e. it will be able to meaningfully respond to open-domain queries (e.g., “Do you like Prague?”).
The model will be based on neural networks (specifically, pre-trained language models). For model training, we plan to use techniques such as transfer learning, adversarial training, iterative training, and combining generative and retrieval-based models. We will also focus on training data augmentation.
The combination of both types of systems will lead to more robust task-oriented chatbots and more useful and controllable chit-chat systems.