Generating digital assistants from source code repositories

Abstract

A machine-learning model is trained using a training set that comprises labeled examples. Each example is a source code repository for an application and each label is a digital assistant for the application. The trained machine-learning model receives a source code repository as input and generates a digital assistant as output. The machine-learning model may be trained using the transformer architecture that is based on attention mechanisms without using recurrence and convolution. Prior to use in training, a collection of source code repositories may be filtered to remove repositories or files that use unknown or unsupported programming languages, that are labeled with incorrect or invalid digital assistant definitions, that include too many or too few files, or that include too many or too few lines of code. A separate machine-learning model may be trained for each programming language, for each digital assistant definition framework, or for each language/framework pair.