HiTyper: A Hybrid Static Type Inference Framework with Neural Prediction Yun Peng Zongjie Li Cuiyun Gao∗ The Chinese University of Hong Kong Harbin Institute of Technology Harbin Institute of Technology Hong Kong, China Shenzhen, China Shenzhen, China
[email protected] [email protected] [email protected] Bowei Gao David Lo Michael Lyu Harbin Institute of Technology Singapore Management University The Chinese University of Hong Kong Shenzhen, China Singapore Hong Kong, China
[email protected] [email protected] [email protected] ABSTRACT also supports type annotations in the Python Enhancement Pro- Type inference for dynamic programming languages is an impor- posals (PEP) [21, 22, 39, 43]. tant yet challenging task. By leveraging the natural language in- Type prediction is a popular task performed by most attempts. formation of existing human annotations, deep neural networks Traditional static type inference techniques [4, 9, 14, 17, 36] and outperform other traditional techniques and become the state-of- type inference tools such as Pytype [34], Pysonar2 [33], and Pyre the-art (SOTA) in this task. However, they are facing some new Infer [31] can predict sound results for the variables with enough challenges, such as fixed type set, type drift, type correctness, and static constraints, e.g., a = 1, but are unable to handle the vari- composite type prediction. ables with few static constraints, e.g. most function arguments. To mitigate the challenges, in this paper, we propose a hybrid On the other hand, dynamic type inference techniques [3, 37] and type inference framework named HiTyper, which integrates static type checkers simulate the workflow of functions and solve types inference into deep learning (DL) models for more accurate type according to input cases and typing rules.