Generative models for dialogue state tracking using RDF

State-of-the-art models assume a dialogue state representation based on concepts (slots) and their associated values. For example the system, at some point in the conversation, may believe the user wants the “price” to be “cheap”. However, real world applications naturally show dependencies between concepts: price and location of a restaurant depend on each other because some part of a city may be more expensive than another. To handle these dependencies, we defined a richer representation based on RDF graphs and trained DST models to generate said graphs.

Resources (code and documents) are available online on GitHub under MIT license.