Predicting Type Annotations for Python using Embeddings from Graph Neural Networks

An intelligent tool for type annotations in Python would increase the productivity of developers. Python is a dynamic programming language, and predicting types using static analysis is difficult. Existing techniques for type prediction use deep learning models originated in the area of Natural Language Processing. These models depend on the quality of embeddings for source code tokens. We compared approaches for pre- training embeddings for source code. Specifically, we compared FastText embeddings to embeddings trained with Graph Neural Networks (GNN). Our experiments showed that GNN embeddings outperformed FastText embeddings on the task of type prediction. Moreover, they seem to encode complementary information since the prediction quality increases when both types of embeddings are used