Semantic Code Classification for Automated Machine Learning

Generating a complete machine learning pipeline from a short description in natural language is a notoriously difficult task because the code is generally much longer than the description and uses a lot of information from different parts of the description. Using intermediate representation might help with this task since specific actions require specific information. In this work, we present a semantic code classification task and a way to represent the machine learning pipeline as a sequence of such semantic classes. Finally, we discuss methods for solving semantic code classification problem on the Natural Language to Machine Learning (NL2ML) dataset.