The current state-of-the-art large language models (LLMs) are effective in generating high-quality text and encapsulating a broad spectrum of world knowledge. However, these models often hallucinate during generation and are not designed to utilize external information sources. To enable requests to the external knowledge bases, also called knowledge grounding, retrieval-augmented LLMs were introduced. For now, their applications have largely involved Open Domain Question Answering, Abstractive Question Answering, and such. In this paper, we broaden the scope of retrieval-augmented LLMs by venturing into a new task - code generation using external entities. For this task, we collect and publish a new dataset for project-level code generation, where the model should reuse functions defined in the project during generation. As we show, existing retrieval-augmented LLMs fail to assign relevance scores between similar entity names, and to mitigate it, they expand entity names with description context and append it to the input. In practice, due to the limited context size they can not accommodate the indefinitely large context of the whole project. To solve this issue, we propose a novel end-to-end trainable architecture with an scalable entity retriever injected directly into the LLM decoder. We demonstrate that our model can outperform common baselines in several scenarios, including project-level code generation, as well as Bash and SQL scripting.
Commit messages are crucial to software development, allowing developers to track changes and collaborate effectively. Despite their utility, most commit messages lack important information since writing high-quality commit messages is tedious and time-consuming. The active research on commit message generation (CMG) has not yet led to wide adoption in practice. We argue that if we could shift the focus from commit message generation to commit message completion and use previous commit history as additional context, we could significantly improve the quality and the personal nature of the resulting commit messages. In this paper, we propose and evaluate both of these novel ideas. Since the existing datasets lack historical data, we collect and share a novel dataset called CommitChronicle, containing 10.7M commits across 20 programming languages. We use this dataset to evaluate the completion setting and the usefulness of the historical context for state-of-the-art CMG models and GPT-3.5-turbo. Our results show that in some contexts, commit message completion shows better results than generation, and that while in general GPT-3.5-turbo performs worse, it shows potential for long and detailed messages. As for the history, the results show that historical information improves the performance of CMG models in the generation task, and the performance of GPT-3.5-turbo in both generation and completion.
Solving a problem with a deep learning model requires researchers to optimize the loss function with a certain optimization method. The research community has developed more than a hundred different optimizers, yet there is scarce data on optimizer performance in various tasks. In particular, none of the benchmarks test the performance of optimizers on source code-related problems. However, existing benchmark data indicates that certain optimizers may be more efficient for particular domains. In this work, we test the performance of various optimizers on deep learning models for source code and find that the choice of an optimizer can have a significant impact on the model quality, with up to two-fold score differences between some of the relatively well-performing optimizers. We also find that RAdam optimizer (and its modification with the Lookahead envelope) is the best optimizer that almost always performs well on the tasks we consider. Our findings show a need for a more extensive study of the optimizers in code-related tasks, and indicate that the ML4SE community should consider using RAdam instead of Adam as the default optimizer for code-related deep learning tasks.
In recent years, researchers have created and introduced a significant number of various code generation models. As human evaluation of every new model version is unfeasible, the community adopted automatic evaluation metrics such as BLEU to approximate the results of human judgement. These metrics originate from the machine translation domain and it is unclear whether they are applicable for the code generation tasks and how well do they agree with the human evaluation on this task. There also are two metrics, CodeBLEU and RUBY, that were developed to estimate the similarity of code and take into account the code properties. However, for these metrics there are hardly any studies on their agreement with the human evaluation. Despite all that, minimal differences in the metric scores are used to claim superiority of some code generation models over the others. In this paper, we present a study on applicability of six metrics -- BLEU, ROUGE-L, METEOR, ChrF, CodeBLEU, RUBY -- for evaluation of the code generation models. We conduct a study on two different code generation datasets and use human annotators to assess the quality of all models run on these datasets. The results indicate that for the CoNaLa dataset of Python one-liners none of the metrics can correctly emulate human judgement on which model is better with $>95\%$ certainty if the difference in model scores is less than 5 points. For the HearthStone dataset, which consists of classes of particular structure, the difference in model scores of at least 2 points is enough to claim the superiority of one model over the other. Using our findings, we derive several recommendations on using metrics to estimate the model performance on the code generation task.
Code clones are pairs of code snippets that implement similar functionality. Clone detection is a fundamental branch of automatic source code comprehension, having many applications in refactoring recommendation, plagiarism detection, and code summarization. A particularly interesting case of clone detection is the detection of semantic clones, i.e., code snippets that have the same functionality but significantly differ in implementation. A promising approach to detecting semantic clones is contrastive learning (CL), a machine learning paradigm popular in computer vision but not yet commonly adopted for code processing. Our work aims to evaluate the most popular CL algorithms combined with three source code representations on two tasks. The first task is code clone detection, which we evaluate on the POJ-104 dataset containing implementations of 104 algorithms. The second task is plagiarism detection. To evaluate the models on this task, we introduce CodeTransformator, a tool for transforming source code. We use it to create a dataset that mimics plagiarised code based on competitive programming solutions. We trained nine models for both tasks and compared them with six existing approaches, including traditional tools and modern pre-trained neural models. The results of our evaluation show that proposed models perform diversely in each task, however the performance of the graph-based models is generally above the others. Among CL algorithms, SimCLR and SwAV lead to better results, while Moco is the most robust approach. Our code and trained models are available at https://doi.org/10.5281/zenodo.6360627, https://doi.org/10.5281/zenodo.5596345.
As researchers and practitioners apply Machine Learning to increasingly more software engineering problems, the approaches they use become more sophisticated. A lot of modern approaches utilize internal code structure in the form of an abstract syntax tree (AST) or its extensions: path-based representation, complex graph combining AST with additional edges. Even though the process of extracting ASTs from code can be done with different parsers, the impact of choosing a parser on the final model quality remains unstudied. Moreover, researchers often omit the exact details of extracting particular code representations. In this work, we evaluate two models, namely Code2Seq and TreeLSTM, in the method name prediction task backed by eight different parsers for the Java language. To unify the process of data preparation with different parsers, we develop SuperParser, a multi-language parser-agnostic library based on PathMiner. SuperParser facilitates the end-to-end creation of datasets suitable for training and evaluation of ML models that work with structural information from source code. Our results demonstrate that trees built by different parsers vary in their structure and content. We then analyze how this diversity affects the models' quality and show that the quality gap between the most and least suitable parsers for both models turns out to be significant. Finally, we discuss other features of the parsers that researchers and practitioners should take into account when selecting a parser along with the impact on the models' quality. The code of SuperParser is publicly available at https://doi.org/10.5281/zenodo.6366591. We also publish Java-norm, the dataset we use to evaluate the models: https://doi.org/10.5281/zenodo.6366599.
Machine Learning for Software Engineering (ML4SE) is an actively growing research area that focuses on methods that help programmers in their work. In order to apply the developed methods in practice, they need to achieve reasonable quality in order to help rather than distract developers. While the development of new approaches to code representation and data collection improves the overall quality of the models, it does not take into account the information that we can get from the project at hand. In this work, we investigate how the model's quality can be improved if we target a specific project. We develop a framework to assess quality improvements that models can get after fine-tuning for the method name prediction task on a particular project. We evaluate three models of different complexity and compare their quality in three settings: trained on a large dataset of Java projects, further fine-tuned on the data from a particular project, and trained from scratch on this data. We show that per-project fine-tuning can greatly improve the models' quality as they capture the project's domain and naming conventions. We open-source the tool we used for data collection, as well as the code to run the experiments: https://zenodo.org/record/6040745.
Integrated Development Environments (IDE) are designed to make users more productive, as well as to make their work more comfortable. To achieve this, a lot of diverse tools are embedded into IDEs, and the developers of IDEs can employ anonymous usage logs to collect the data about how they are being used to improve them. A particularly important component that this can be applied to is code completion, since improving code completion using statistical learning techniques is a well-established research area. In this work, we propose an approach for collecting completion usage logs from the users in an IDE and using them to train a machine learning based model for ranking completion candidates. We developed a set of features that describe completion candidates and their context, and deployed their anonymized collection in the Early Access Program of IntelliJ-based IDEs. We used the logs to collect a dataset of code completions from users, and employed it to train a ranking CatBoost model. Then, we evaluated it in two settings: on a held-out set of the collected completions and in a separate A/B test on two different groups of users in the IDE. Our evaluation shows that using a simple ranking model trained on the past user behavior logs significantly improved code completion experience. Compared to the default heuristics-based ranking, our model demonstrated a decrease in the number of typing actions necessary to perform the completion in the IDE from 2.073 to 1.832. The approach adheres to privacy requirements and legal constraints, since it does not require collecting personal information, performing all the necessary anonymization on the client's side. Importantly, it can be improved continuously: implementing new features, collecting new data, and evaluating new models - this way, we have been using it in production since the end of 2020.
A recent study by Ahmed and Devanbu reported that using a corpus of code written in multilingual datasets to fine-tune multilingual Pre-trained Language Models (PLMs) achieves higher performance as opposed to using a corpus of code written in just one programming language. However, no analysis was made with respect to fine-tuning monolingual PLMs. Furthermore, some programming languages are inherently different and code written in one language usually cannot be interchanged with the others, i.e., Ruby and Java code possess very different structure. To better understand how monolingual and multilingual PLMs affect different programming languages, we investigate 1) the performance of PLMs on Ruby for two popular Software Engineering tasks: Code Summarization and Code Search, 2) the strategy (to select programming languages) that works well on fine-tuning multilingual PLMs for Ruby, and 3) the performance of the fine-tuned PLMs on Ruby given different code lengths. In this work, we analyze over a hundred of pre-trained and fine-tuned models. Our results show that 1) multilingual PLMs have a lower Performance-to-Time Ratio (the BLEU, METEOR, or MRR scores over the fine-tuning duration) as compared to monolingual PLMs, 2) our proposed strategy to select target programming languages to fine-tune multilingual PLMs is effective: it reduces the time to fine-tune yet achieves higher performance in Code Summarization and Code Search tasks, and 3) our proposed strategy consistently shows good performance on different code lengths.
The task of finding the best developer to fix a bug is called bug triage. Most of the existing approaches consider the bug triage task as a classification problem, however, classification is not appropriate when the sets of classes change over time (as developers often do in a project). Furthermore, to the best of our knowledge, all the existing models use textual sources of information, i.e., bug descriptions, which are not always available. In this work, we explore the applicability of existing solutions for the bug triage problem when stack traces are used as the main data source of bug reports. Additionally, we reformulate this task as a ranking problem and propose new deep learning models to solve it. The models are based on a bidirectional recurrent neural network with attention and on a convolutional neural network, with the weights of the models optimized using a ranking loss function. To improve the quality of ranking, we propose using additional information from version control system annotations. Two approaches are proposed for extracting features from annotations: manual and using an additional neural network. To evaluate our models, we collected two datasets of real-world stack traces. Our experiments show that the proposed models outperform existing models adapted to handle stack traces. To facilitate further research in this area, we publish the source code of our models and one of the collected datasets.