CAT

Natural language interfaces for simple problems like booking a flight are becoming ubiquitous. While database transactions are often the backbone for such business processes, it still comes at cost to develop a conversational agent for a given OLTP application (e.g., allowing users to buy a flight ticket), since it requires both immense amounts of training data and NLP expertise.

In this project, we aim at automating the creation of conversational agents for a given OLTP application with only minimal manual overhead using a system called CAT. For a given relational database and a set of transactions, CAT synthesizes the required training data with weak supervision to train a state-of-the-art conversational model that allows users to interact with the database. A major difference to existing conversational agents is that the trained agents are data-aware. In particular, we make the decision which information should be requested from the user based on the actual data distribution, i.e., by choosing informative attributes which the user likely knows resulting in shorter overall dialogues.