Do Contexts Help in Phrase-Based, Statistical Source Code Migration?

摘要

Prior research showed that to migrate Java code to C# by directly applying phrase-based statistical machine translation (SMT) on the lexemes of source code produces much semantically incorrect code. In this work, we conduct empirical studies on several open-source projects to investigate the use of well-defined semantics in programming languages to guide the translation process in SMT. We have investigated five types of features forming the contexts involving the (semantic) relations among code tokens including occurrence association among code tokens, data and control dependencies among program entities, visibility constraints of entities, and the consistency in declarations and accesses of variables, fields and methods. We use the Direct Maximum Entropy (DME) approach for feature integration. Our empirical results show that as individual features added to the baseline SMT model, token association and data dependencies contribute much with highest relative improvement in semantic correctness of up to 18.3% and 18.5%, respectively. The integration of three feature types (token association, data dependencies, and visibility) into the baseline model has highest relative improvement with up to 26.4% improvement in semantic correctness. Generally, 43.5-80.7% of the total translated methods are semantically correct. Our results show a good direction of using SMT with semantic features at different levels of abstraction to improve its accuracy.