Data-driven program translation has been recently the focus of several lines of research. A common and robust strategy is supervised learning. However, there is typically a lack of parallel training data, i.e., pairs of code snippets in the source and target language. While many data augmentation techniques exist in the domain of natural language processing, they cannot be easily adapted to tackle code translation due to the unique restrictions of programming languages. In this paper, we develop a novel rule-based augmentation approach tailored for code translation data, and a novel retrieval-based approach that combines code samples from unorganized big code repositories to obtain new training data. Both approaches are language-independent. We perform an extensive empirical evaluation on existing Java-C#-benchmarks showing that our method improves the accuracy of state-of-the-art supervised translation techniques by up to 35%.

Coding style has direct impact on code comprehension. Automatically transferring code style to user's preference or consistency can facilitate project cooperation and maintenance, as well as maximize the value of open-source code. Existing work on automating code stylization is either limited to code formatting or requires human supervision in pre-defining style checking and transformation rules. In this paper, we present unsupervised methods to assist automatic code style transfer for arbitrary code styles. The main idea is to leverage Big Code database to learn style and content embedding separately to generate or retrieve a piece of code with the same functionality and the desired target style. We carefully encode style and content features, so that a style embedding can be learned from arbitrary code. We explored the capabilities of novel attention-based style generation models and meta-learning and implemented our ideas in DUETCS. We complement the learning-based approach with a retrieval mode, which uses the same embeddings to directly search for the desired piece of code in Big Code. Our experiments show that DUETCS captures more style aspects than existing baselines.

Cross-language code retrieval is necessary in many real-world scenarios. A major application is program translation, e.g., porting codebases from an obsolete or deprecated language to a modern one or re-implementing existing projects in one’s preferred programming language. Existing approaches based on the translation model require large amounts of training data and extra information or neglects significant characteristics of programs. Leveraging cross-language code retrieval to assist automatic program translation can make use of Big Code. However, existing code retrieval systems have the barrier to finding the translation with only the features of the input program as the query. In this paper, we present BigPT for interactive cross-language retrieval from Big Code only based on raw code and reusing the retrieved code to assist program translation. We build on existing work on cross-language code representation and propose a novel predictive transformation model based on auto-encoders. The model is trained on Big Code to generate a target-language representation, which will be used as the query to retrieve the most relevant translations for a given program. Our query representation enables the user to easily update and correct the returned results to improve the retrieval process. Our experiments show that BigPT outperforms state-of-the-art baselines in terms of program accuracy. Using our novel querying and retrieving mechanism, BigPT can be scaled to the large dataset and efficiently retrieve the translation.