With the vigorous rise of software engineering practices,and the thriving development of open-source communities,deep learning-based program synthesis has emerged as a focal point of interest in both academia and industry.This field encompasses a range of disciplines including software engineering,deep learning,data mining,natural language processing,and programming languages.Deep learning-based program synthesis,namely intelligent program synthesis,utilizes deep learning techniques to extract knowledge from vast program repositories,with the goal of creating smart tools that improve the quality and productivity of computer programming.In contrast to traditional synthesis methods reliant on heuristic rules or expert systems,program intelligent synthesis has swiftly gained prominence due to its highly scalable and self-optimizing characteristics,becoming a research focus on both software engineering and artificial intelligence domains.The rapid advancement of pre-training techniques has led to the increasing adoption of large-scale language models in program synthesis,propelling continuous advancements in this domain.For example,GPT-4 has demonstrated human-comparable performance on platforms like LeetCode,while DeepMind's AlphaCode addresses challenges in natural language competitive programming.Simultaneously,the industry has introduced a series of AI programming assistants such as Copilot,Comate,and CodeWhisperer,significantly enhancing development efficiency and drastically reducing the learning curve in programming,thereby enabling broader participation in software development.To foster deeper research and widespread application in this field,this paper systematically explores the latest research progress in program intelligent synthesis from various perspectives.It comprehensively discusses aspects such as user intent understanding,program comprehension,model training,model testing,and evaluation,with detailed subdivisions.User intent understanding aims to locate and understand user intentions by integrating contextual semantics and knowledge swiftly and accurately.The paper introduces methods for understanding users from different angles,including input-output pairs,natural language,programs,and visual aspects.Program comprehension analyzes and extracts critical information from programs at various abstraction levels and perspectives,transforming it into forms understandable by computers.This paper presents program comprehension methods based on text sequences,tree structures,and graph structures.Model training uses this information to generate new code,while model testing and evaluation verify and optimize the quality and performance of generated code.The paper also examines challenges such as uneven dataset quality,low efficiency in user intent under-standing and program comprehension,as well as issues regarding model interpretability and robustness.Furthermore,the paper anticipates future trends,including higher-quality datasets,more efficient methods for user intent understanding and program comprehension,more robust model architectures,and improved application of these technologies in practical industrial settings.This research not only aids the academic community in comprehensively understanding the latest developments in the field of intelligent program synthesis but also assists software developers in quickly mastering relevant technologies and strategies to meet industrial demands.Through continuous exploration and innovation,intelligent program synthesis is poised to achieve greater breakthroughs in the future,driving innovation and development across the entire software engineering domain.The integration of these advancements promises to revolutionize software engineering practices,ushering in an era of enhanced efficiency and creativity in programming and development workflows.
万方数据
维普
