PSR-SQUARES:基于程序空间约简器的SQL逆向合成系统

PSR-SQUARES:SQL reverse synthesis system based on program space reducer

针对SQUARES程序空间增长过快,导致程序合成效率偏低的问题,在SQUARES的基础上,增加了以深度神经网络为核心的程序空间约简器,将给定的<被查询表,查询结果>示例表示成二维张量,作为深度神经网络的输入,网络的输出是关于目标SQL语句合成规则的相关性标记向量。约简器根据神经网络的输出结果,采用末N位淘汰策略,删除与目标SQL语句相关性弱的合成规则,以减少候选SQL语句的生成和验证,提升系统合成效率。对约简器中深度神经网络的结构设计、训练样本集的生成方法和网络训练过程进行了详细描述。同时将PSR-SQUARES与当前有代表性SQL逆向合成系统进行实验对比,实验结果表明,PSR-SQUARES的综合性能不同程度地优于其他合成系统,平均合成时间由SQUARES的251 s降低至130 s,目标程序合成成功率由80%提升至89%。

In order to address the issue of rapid growth of program space in SQUARES,which led to low efficiency in program synthesis,a program space reducer based on deep neural network(DNN)was introduced into the SQUARES framework.A given<Queried tables,Query result>pair was represented as a 2D tensor which was used as input for a DNN.And the output of the DNN was the relevance vector of the target SQL statement synthesis rules.Based on the output of the DNN,the last N rules with weak correlation to the target SQL statement were eliminated,thereby shrinking the program search space and improving the system synthesis efficiency.The architecture of DNN,the method of generating training datasets,and the training process of DNN were described in detail.Furthermore,experimental comparisons between PSR-SQUARES and other representative SQL reverse synthesis systems were conducted.The results show that the overall performance of PSR-SQUARES is superior to other synthesis systems to varying degrees,with the average synthesis time reduced from 251 s in SQUARES to 130 s and the target program synthesis success rate increased from 80%to 89%.