Exploring Paraphrasing Techniques on Formal Language for Generating Semantics Preserving Source Code Transformations
Aviel J. Stein, Levi Kapllani, Spiros Mancoridis, Rachel Greenstadt · 2020
Automatically identifying and generating equivalent semantic content to a word, phrase, or sentence is an important part of natural language processing (NLP). The research done so far in paraphrases in NLP has been focused exclusively on textual data, but has significant potential if it is applied to formal languages like source code. In this paper, we present a novel technique for generating source code transformations via the use of paraphrases. We explore how to extract and validate source code paraphrases. The transformations can be used for stylometry tasks and processes like refactoring. A machine learning method of identifying valid transformations has the advantage of avoiding the generation of transformations by hand and is more likely to have more valid transformations. Our dataset is comprised by 27,300 C++ source code files, consisting of 273topics each with 10 parallel files. This generates approximately152,000 paraphrases. Of these paraphrases, 11% yield valid code transformations. We then train a random forest classifier that can identify valid transformations with 83% accuracy. In this paper we also discuss some of the observed relationships betweenlinked paraphrase transformations. We depict the relationshipsthat emerge between alternative equivalent code transformationsin a graph formalism.