Scientists from Salesforce Research and Chinese University of Hong Kong have released Photon, a live prototype of a natural language interface to complex relational databases.

The team used deep-learning to design Photon following two core principles: intelligence and robustness.

It adopts a modular architecture consisting of a neural semantic parser, a human-in-the-loop question corrector, a DB engine and a response generator.

The team expects the interface to correctly interpret a diverse set of natural language questions, while avoiding unreliable guesses for noisy input.

At the core of Photon is a semantic parser. It helps in mapping a natural language user input to an executable SQL query. Further, Photon is claimed to adopt a cross-DB semantic parsing model that realizes this mapping for a large number of DBs, including DBs it has never been trained on.

Also, Photon accepts both natural language questions and well-formed SQL queries as input. The idea is to automatically detect the input type and execute the input immediately in case it’s a valid SQL query.

The current Photon system is still a prototype, with limited functions.

The team has plans to add more features to Photon, including voice input, auto-completion, and visualization of the output when appropriate.