面向App软件开发的API通用表示模型预训练方法及其应用研究

API的使用是提高App产品开发效率、提升产品质量的重要途径。如何合理使用API资源一直以来都是软件工程领域重要的研究课题。研究表明，利用机器学习方法从大规模数据中获取API使用知识是解决该问题的有效途径。近年来，BERT、GPT-3等模型的出现引发深度学习模式的新浪潮，通过超大规模的模型预训练，能够实现针对不同问题，以“低花费”获取“高质量”的可用模型。本项目拟基于App商店中超大规模代码数据，研究API通用表示模型的预训练方法，利用App特征信息引导预训练过程，克服代码数据训练中的难点问题。项目预期实现两个目标：1、提供通用的API表示预训练模型，阐明其预训练方法并提供相关数据集；2、针对API推荐，API误用预警以及API迁移分析三个具体任务进行模型微调，在检验API表示模型预训练效果的同时，解决App开发中的实际API使用问题。

The API (Application Programming Interfaces) is important for improving the efficiency and quality of App development, and how to use the APIs properly is always a hot research question in Software Engineering. With the development of machine learning, many studies show that it is an efficient way for solving this problem. Recently, BERT and GPT-3 trigger a new wave of deep learning models: after the large-scale pre-training of the model, it can be fine tuned with just one additional output layer to create state-of-the-art models for different tasks, which means we can spend lower cost to gain higher quality models. Following this idea, this project aims to pre-training a general representation model of APIs based on the large-scale of data in App stores. Considering the Apps take the development of “features” as the core, we want to introduce the information of App features to tackle the difficulties caused by the characteristics of codes in the training process (such as code divisions, input embedding, and pre-training task), so that we can gain the pre-trained models with lower cost. The project is intended to achieve two goals: 1. providing a usable pre-trained model of API presentations, and illustrating the training method as well as the training data; 2. fine-tuning the model for three specific tasks, including API recommendation, API misuse detection and API migration analysis, in this way, we could not only validate the performance of the pre-trained model, but also help developers solving the practice problems of using API resource in the development of Apps.