Access to information is increasingly conversational in nature. Task-oriented chat systems help users achieve goals, at helpdesks, while shopping, etc. To be engaging, modern chat systems are computationally intensive and data hungry, requiring large-scale language-specific, domain-specific and task-specific training data. Our central aim is to enable training, deployment and adaptation of effective and efficient, task-oriented chat-based systems that facilitate end users to interact with widely used services in a natural manner.
The core research questions underlying LESSEN are:
· How can we build and adapt chat-based conversational agents in a computationally and data efficient manner?
· How can we automatically generate, simulate and transfer training data so as to produce engaging chat-based conversations across many tasks and domains?
· How can we do all this in a safe and transparent manner, with provisions for explainability and data provenance?
LESSEN’s target outcomes are the following. First, algorithms (1) for automatically improving neural architectures for chat-based agents so as to achieve efficiency gains at training and inference time; (2) for making more efficient use of conversational training data; (3) for domain adaptation for conversational agents; and (4) for augmenting conversational training data. Second, methods (1) for ensuring safe and privacy-preserving conversational agents, and (2) for providing transparency for conversational agents. Third, societal impact (1) in a direct manner by directly influencing conversational agent development at industrial partners in the consortium, and (2) in an indirect manner by engaging with governmental and non-governmental stakeholders and with a branch organization to reach out to an entire sector (retail).
The societal breakthrough that LESSEN seeks to create is to democratize conversational AI technology. Our consortium is keen to help bring modern conversational AI technology within reach of a diverse set of industrial, governmental, and non-governmental stakeholders.
